ai agents

Finding the Agentic Sweet Spot

<! DOCTYPE html><html xmlns="http://www.

Finding the Agentic Sweet Spot
September 18, 20265 min readai agents
<!DOCTYPE html><html xmlns="http://www.w3.org/1999/xhtml" lang="" xml:lang=""><head><title>Agentic-Sweet-Spot</title><meta http-equiv="Content-Type" content="text/html; charset=UTF-8"/><meta name="generator" content="pdftohtml 0.36"/><meta name="author" content="sashsarangi"/><meta name="keywords" content="DAHVd3tRZHI,BADKjKpPWtI"/><meta name="date" content="2026-09-17T23:50:03+00:00"/><style type="text/css"><!--.xflip { -moz-transform: scaleX(-1); -webkit-transform: scaleX(-1); -o-transform: scaleX(-1); transform: scaleX(-1); filter: fliph;}.yflip { -moz-transform: scaleY(-1); -webkit-transform: scaleY(-1); -o-transform: scaleY(-1); transform: scaleY(-1); filter: flipv;}.xyflip { -moz-transform: scaleX(-1) scaleY(-1); -webkit-transform: scaleX(-1) scaleY(-1); -o-transform: scaleX(-1) scaleY(-1); transform: scaleX(-1) scaleY(-1); filter: fliph + flipv;}--></style></head><body bgcolor="#A0A0A0" vlink="blue" link="blue"><!-- Page 1 --><a name="1"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft10{font-size:52px;font-family:AAAAAA+OpenSauceOne;color:#ffffff;} .ft11{font-size:36px;font-family:AAAAAA+OpenSauceOne;color:#d8d6d6;}--></style><div id="page1-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index001.png" alt="background image"/><p style="position:absolute;top:223px;left:78px;white-space:nowrap" class="ft10"><b>Finding&#160;the&#160;Agentic&#160;Sweet&#160;Spot</b></p><p style="position:absolute;top:295px;left:78px;white-space:nowrap" class="ft11"><b>Matching&#160;Architecture&#160;to&#160;Enterprise&#160;Constraints</b></p></div><!-- Page 2 --><a name="2"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft22{font-size:49px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft23{font-size:18px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft24{font-size:18px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft25{font-size:25px;font-family:BAAAAA+OpenSauceOne;color:#535353;} .ft26{font-size:24px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft27{font-size:15px;font-family:BAAAAA+OpenSauceOne;color:#535353;} .ft28{font-size:11px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft29{font-size:27px;font-family:BAAAAA+OpenSauceOne;color:#1a1a1a;} .ft210{font-size:15px;font-family:AAAAAA+OpenSauceOne;color:#535353;} .ft211{font-size:18px;line-height:24px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft212{font-size:18px;line-height:24px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft213{font-size:15px;line-height:20px;font-family:AAAAAA+OpenSauceOne;color:#535353;}--></style><div id="page2-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index002.png" alt="background image"/><p style="position:absolute;top:38px;left:70px;white-space:nowrap" class="ft22">Executive&#160;Summary</p><p style="position:absolute;top:178px;left:305px;white-space:nowrap" class="ft211"><b>Enterprise&#160;agentic&#160;systems&#160;rarely&#160;improve&#160;monotonically&#160;as<br/>more&#160;agents,&#160;tools,&#160;and&#160;reasoning&#160;stages&#160;are&#160;added.&#160;In&#160;this<br/>white&#160;paper,&#160;we&#160;present&#160;a&#160;comparative&#160;study&#160;of&#160;agentic<br/>architectures&#160;of&#160;varying&#160;complexity&#160;for&#160;predicting&#160;currency<br/>volatility.&#160;In&#160;this&#160;experiment,&#160;a&#160;simple&#160;hybrid&#160;combining&#160;a<br/>traditional&#160;quantitative&#160;model&#160;with&#160;a&#160;low-cost&#160;LLM&#160;significantly<br/>improved&#160;the&#160;results,&#160;while&#160;full&#160;orchestration&#160;of&#160;LLM-based<br/>agents&#160;was&#160;substantially&#160;more&#160;expensive&#160;and&#160;less&#160;accurate.<br/>The&#160;result&#160;is&#160;not&#160;a&#160;universal&#160;ranking&#160;of&#160;models.&#160;It&#160;is&#160;a&#160;design<br/>lesson:&#160;the&#160;most&#160;efficient&#160;agentic&#160;architecture&#160;depends&#160;on&#160;the<br/>problem,&#160;the&#160;information&#160;available&#160;at&#160;decision&#160;time,&#160;and&#160;the<br/>appropriate&#160;division&#160;of&#160;tasks&#160;between&#160;traditional&#160;models&#160;and<br/>LLMs.</b></p><p style="position:absolute;top:525px;left:305px;white-space:nowrap" class="ft24">The&#160;study&#160;led&#160;to&#160;the&#160;following&#160;lessons:</p><p style="position:absolute;top:574px;left:335px;white-space:nowrap" class="ft23"><b>An&#160;agentic&#160;sweet&#160;spot&#160;is&#160;plausible:&#160;</b></p><p style="position:absolute;top:574px;left:644px;white-space:nowrap" class="ft24">The&#160;observed</p><p style="position:absolute;top:599px;left:335px;white-space:nowrap" class="ft212">configurations&#160;formed&#160;a&#160;practical&#160;cost–accuracy&#160;trade-off.<br/>The&#160;light&#160;hybrid&#160;achieved&#160;better&#160;accuracy&#160;at&#160;an&#160;estimated<br/>cost&#160;of&#160;$0.015&#160;per&#160;case.&#160;The&#160;orchestrated&#160;hybrid&#160;delivered<br/>lower&#160;accuracy&#160;at&#160;a&#160;higher&#160;estimated&#160;cost&#160;of&#160;$0.093&#160;per<br/>case.&#160;In&#160;this&#160;experiment,&#160;additional&#160;orchestration&#160;did&#160;not<br/>produce&#160;better&#160;measured&#160;performance.</p><p style="position:absolute;top:772px;left:335px;white-space:nowrap" class="ft23"><b>Domain&#160;structure&#160;determines&#160;architecture:&#160;</b></p><p style="position:absolute;top:772px;left:711px;white-space:nowrap" class="ft24">The&#160;system</p><p style="position:absolute;top:797px;left:335px;white-space:nowrap" class="ft212">forecast&#160;one-hour-ahead&#160;USDJPY&#160;realized&#160;volatility&#160;around<br/>scheduled&#160;macroeconomic&#160;events.&#160;Quantitative&#160;finance<br/>provided&#160;natural&#160;numerical-model&#160;anchors,&#160;a&#160;known&#160;event<br/>calendar,&#160;and&#160;a&#160;bounded&#160;action&#160;space.&#160;A&#160;different&#160;problem&#160;-<br/>such&#160;as&#160;multi-document&#160;reconciliation&#160;or&#160;policy&#160;interpretation<br/>-&#160;could&#160;justify&#160;a&#160;richer&#160;agentic&#160;design.</p><p style="position:absolute;top:970px;left:335px;white-space:nowrap" class="ft23"><b>Separate&#160;calculation&#160;from&#160;context</b></p><p style="position:absolute;top:970px;left:635px;white-space:nowrap" class="ft24">:&#160;Quant&#160;libraries&#160;should</p><p style="position:absolute;top:995px;left:335px;white-space:nowrap" class="ft212">calculate&#160;returns,&#160;realized&#160;volatility,&#160;GARCH&#160;forecasts,<br/>losses,&#160;and&#160;statistical&#160;tests.&#160;LLMs&#160;should&#160;interpret&#160;contextual<br/>evidence,&#160;identify&#160;relevant&#160;regimes,&#160;and&#160;select&#160;among&#160;pre-<br/>approved&#160;actions.&#160;LLMs&#160;should&#160;not&#160;replace&#160;arithmetic&#160;that&#160;is<br/>already&#160;auditable&#160;and&#160;robust.</p><p style="position:absolute;top:650px;left:102px;white-space:nowrap" class="ft25"><b>Budget</b></p><p style="position:absolute;top:782px;left:104px;white-space:nowrap" class="ft25"><b>Quality</b></p><p style="position:absolute;top:912px;left:78px;white-space:nowrap" class="ft25"><b>Compliance</b></p><p style="position:absolute;top:106px;left:70px;white-space:nowrap" class="ft26">Choosing&#160;the&#160;Right&#160;Agentic&#160;Architecture&#160;for&#160;Enterprise&#160;Tasks</p><p style="position:absolute;top:690px;left:50px;white-space:nowrap" class="ft27"><b>Operate&#160;within&#160;the&#160;allocated</b></p><p style="position:absolute;top:710px;left:125px;white-space:nowrap" class="ft27"><b>budget.</b></p><p style="position:absolute;top:819px;left:22px;white-space:nowrap" class="ft27"><b>Meet&#160;defined&#160;quality&#160;and&#160;accuracy</b></p><p style="position:absolute;top:839px;left:100px;white-space:nowrap" class="ft27"><b>benchmarks.</b></p><p style="position:absolute;top:950px;left:51px;white-space:nowrap" class="ft27"><b>Adhere&#160;to&#160;governance&#160;and</b></p><p style="position:absolute;top:970px;left:65px;white-space:nowrap" class="ft27"><b>compliance&#160;guardrails.</b></p><p style="position:absolute;top:1135px;left:873px;white-space:nowrap" class="ft28">1</p><p style="position:absolute;top:548px;left:37px;white-space:nowrap" class="ft29"><b>Enterprise&#160;Design</b></p><p style="position:absolute;top:585px;left:77px;white-space:nowrap" class="ft29"><b>Constraints</b></p><p style="position:absolute;top:232px;left:26px;white-space:nowrap" class="ft27"><b>About&#160;EMAlpha.&#160;</b></p><p style="position:absolute;top:232px;left:145px;white-space:nowrap" class="ft210">EMAlpha&#160;designs</p><p style="position:absolute;top:252px;left:26px;white-space:nowrap" class="ft213">domain-specific&#160;agents&#160;for&#160;industry<br/>workflows.&#160;Its&#160;pre-trade&#160;agents<br/>provides&#160;real-time&#160;market&#160;updates<br/>and&#160;contextual&#160;intelligence&#160;for<br/>financial&#160;applications,&#160;illustrating<br/>how&#160;the&#160;calculation-versus-context<br/>principles&#160;in&#160;this&#160;paper&#160;can&#160;be<br/>applied&#160;in&#160;production&#160;settings.</p></div><!-- Page 3 --><a name="3"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft314{font-size:18px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft315{font-size:15px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft316{font-size:14px;font-family:CAAAAA+TTHoves;color:#ffffff;} .ft317{font-size:14px;font-family:DAAAAA+TTHoves;color:#000000;} .ft318{font-size:30px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft319{font-size:18px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft320{font-size:11px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft321{font-size:15px;font-family:EAAAAA+OpenSauceOne;color:#1a1a1a;} .ft322{font-size:15px;font-family:FAAAAA+OpenSauceOne;color:#1a1a1a;} .ft323{font-size:15px;line-height:20px;font-family:BAAAAA+OpenSauceOne;color:#000000;} .ft324{font-size:14px;line-height:19px;font-family:CAAAAA+TTHoves;color:#ffffff;} .ft325{font-size:15px;line-height:20px;font-family:FAAAAA+OpenSauceOne;color:#1a1a1a;}--></style><div id="page3-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index003.png" alt="background image"/><p style="position:absolute;top:203px;left:25px;white-space:nowrap" class="ft314"><b>Finding&#160;the&#160;Right&#160;Architecture</b></p><p style="position:absolute;top:246px;left:25px;white-space:nowrap" class="ft323">In&#160;finance,&#160;the&#160;efficient-frontier&#160;concept&#160;describes&#160;the&#160;best&#160;attainable&#160;trade-offs&#160;between&#160;risk&#160;and&#160;expected&#160;return.<br/>For&#160;a&#160;given&#160;level&#160;of&#160;risk,&#160;different&#160;asset&#160;mixes&#160;can&#160;produce&#160;different&#160;expected&#160;returns.&#160;The&#160;efficient&#160;frontier<br/>represents&#160;the&#160;best&#160;attainable&#160;risk–return&#160;combinations..&#160;We&#160;use&#160;a&#160;similar&#160;lens&#160;for&#160;agentic&#160;architecture.&#160;Enterprise<br/>systems&#160;face&#160;competing&#160;constraints&#160;from&#160;budgets,&#160;business&#160;owners,&#160;risk&#160;functions,&#160;and&#160;compliance&#160;teams.&#160;A<br/>common&#160;assumption&#160;is&#160;that&#160;choosing&#160;lower-cost&#160;models&#160;necessarily&#160;sacrifices&#160;quality&#160;or&#160;accuracy.&#160;Our&#160;experience<br/>combining&#160;business&#160;outcomes&#160;with&#160;data&#160;science&#160;led&#160;us&#160;to&#160;question&#160;that&#160;simplification.&#160;This&#160;experiment&#160;was&#160;designed<br/>as&#160;a&#160;concrete&#160;test&#160;of&#160;whether&#160;a&#160;more&#160;expensive&#160;or&#160;more&#160;elaborate&#160;architecture&#160;actually&#160;delivers&#160;enough&#160;measurable<br/>value&#160;to&#160;justify&#160;its&#160;additional&#160;cost&#160;and&#160;complexity.</p><p style="position:absolute;top:429px;left:25px;white-space:nowrap" class="ft323">We&#160;tested&#160;the&#160;idea&#160;on&#160;one&#160;specific&#160;forecasting&#160;problem:&#160;one-hour-ahead&#160;volatility&#160;for&#160;the&#160;USDJPY&#160;currency&#160;pair.<br/>USDJPY&#160;is&#160;a&#160;useful&#160;test&#160;case&#160;because&#160;it&#160;is&#160;sensitive&#160;to&#160;monetary&#160;policy,&#160;central-bank&#160;intervention&#160;risk,&#160;evolving&#160;policy<br/>regimes,&#160;and&#160;shifts&#160;in&#160;global&#160;risk&#160;sentiment.&#160;The&#160;experiment&#160;translated&#160;the&#160;proposed&#160;“AI&#160;efficient&#160;frontier”&#160;into&#160;four<br/>comparable&#160;architecture&#160;designs.&#160;A0&#160;was&#160;a&#160;rolling&#160;GARCH&#160;volatility&#160;forecast,&#160;a&#160;conventional&#160;quantitative&#160;baseline<br/>used&#160;for&#160;decades&#160;with&#160;varying&#160;degrees&#160;of&#160;success.&#160;A1&#160;added&#160;one&#160;constrained&#160;language-model&#160;call&#160;that&#160;received&#160;a<br/>compact&#160;market&#160;and&#160;event&#160;packet&#160;and&#160;selected&#160;a&#160;multiplier&#160;from&#160;a&#160;fixed&#160;grid.&#160;A2&#160;used&#160;four&#160;specialists,&#160;synthesis,&#160;and<br/>a&#160;critic&#160;before&#160;selecting&#160;the&#160;same&#160;bounded&#160;multiplier.&#160;A3&#160;removed&#160;the&#160;numerical&#160;anchor&#160;and&#160;asked&#160;a&#160;language<br/>model&#160;to&#160;forecast&#160;volatility&#160;directly.</p><p style="position:absolute;top:645px;left:84px;white-space:nowrap" class="ft316"><b>Architecture&#160;Family</b></p><p style="position:absolute;top:645px;left:308px;white-space:nowrap" class="ft316"><b>Role&#160;of&#160;LLM</b></p><p style="position:absolute;top:645px;left:520px;white-space:nowrap" class="ft316"><b>Error</b></p><p style="position:absolute;top:636px;left:676px;white-space:nowrap" class="ft324"><b>Estimated<br/>cost/case</b></p><p style="position:absolute;top:716px;left:79px;white-space:nowrap" class="ft317">A0&#160;-&#160;Pure&#160;Quant&#160;Model</p><p style="position:absolute;top:707px;left:277px;white-space:nowrap" class="ft317">No&#160;LLM,&#160;just&#160;traditional</p><p style="position:absolute;top:726px;left:305px;white-space:nowrap" class="ft317">quant&#160;models</p><p style="position:absolute;top:716px;left:520px;white-space:nowrap" class="ft317">3.546</p><p style="position:absolute;top:716px;left:686px;white-space:nowrap" class="ft317">$0.000</p><p style="position:absolute;top:787px;left:99px;white-space:nowrap" class="ft317">A1&#160;-&#160;Light&#160;Hybrid&#160;</p><p style="position:absolute;top:778px;left:280px;white-space:nowrap" class="ft317">Light&#160;touch&#160;LLMs&#160;that</p><p style="position:absolute;top:797px;left:264px;white-space:nowrap" class="ft317">slightly&#160;tweak&#160;quant&#160;model&#160;</p><p style="position:absolute;top:787px;left:519px;white-space:nowrap" class="ft317">2.049</p><p style="position:absolute;top:787px;left:689px;white-space:nowrap" class="ft317">$0.015</p><p style="position:absolute;top:858px;left:80px;white-space:nowrap" class="ft317">A2&#160;-&#160;Orchestrated&#160;LLM</p><p style="position:absolute;top:849px;left:283px;white-space:nowrap" class="ft317">LLMs&#160;as&#160;specialists&#160;+</p><p style="position:absolute;top:868px;left:296px;white-space:nowrap" class="ft317">synthesis&#160;+&#160;critic</p><p style="position:absolute;top:858px;left:520px;white-space:nowrap" class="ft317">2.755</p><p style="position:absolute;top:858px;left:687px;white-space:nowrap" class="ft317">$0.093</p><p style="position:absolute;top:929px;left:107px;white-space:nowrap" class="ft317">A3&#160;-&#160;Pure&#160;LLM</p><p style="position:absolute;top:929px;left:265px;white-space:nowrap" class="ft317">Direct&#160;forecast&#160;using&#160;LLMs</p><p style="position:absolute;top:929px;left:520px;white-space:nowrap" class="ft317">2.884</p><p style="position:absolute;top:929px;left:689px;white-space:nowrap" class="ft317">$0.018</p><p style="position:absolute;top:80px;left:28px;white-space:nowrap" class="ft318"><b>Matching&#160;Agentic&#160;Architecture&#160;to&#160;Enterprise&#160;Requirements</b></p><p style="position:absolute;top:126px;left:28px;white-space:nowrap" class="ft319">Lessons&#160;from&#160;a&#160;quant–LLM&#160;experiment&#160;in&#160;architecture,&#160;cost,&#160;and&#160;quality</p><p style="position:absolute;top:1135px;left:870px;white-space:nowrap" class="ft320">2</p><p style="position:absolute;top:983px;left:32px;white-space:nowrap" class="ft321"><i><b>Table&#160;1.&#160;Primary&#160;comparison&#160;of&#160;four&#160;architecture&#160;families&#160;for&#160;one-hour-ahead&#160;USDJPY&#160;volatility&#160;forecasting.</b></i></p><p style="position:absolute;top:983px;left:801px;white-space:nowrap" class="ft322"><i>&#160;The</i></p><p style="position:absolute;top:1003px;left:32px;white-space:nowrap" class="ft325"><i>study&#160;covered&#160;136&#160;separate&#160;forecast&#160;cases,&#160;including&#160;officially&#160;scheduled&#160;event&#160;periods&#160;and&#160;quieter&#160;control&#160;periods.<br/>Each&#160;LLM-based&#160;architecture&#160;was&#160;run&#160;three&#160;times&#160;on&#160;the&#160;same&#160;cases&#160;to&#160;distinguish&#160;repeatable&#160;performance&#160;patterns<br/>from&#160;nondeterministic&#160;model&#160;variation.&#160;Forecast&#160;error&#160;was&#160;calculated&#160;using&#160;the&#160;quasi-likelihood&#160;loss&#160;metric&#160;described<br/>earlier.&#160;Estimated&#160;costs&#160;represent&#160;LLM&#160;inference&#160;costs&#160;based&#160;on&#160;logged&#160;tokens&#160;and&#160;frozen&#160;catalog&#160;prices;&#160;they<br/>exclude&#160;engineering,&#160;infrastructure,&#160;storage,&#160;and&#160;local&#160;quantitative&#160;computation.</i></p></div><!-- Page 4 --><a name="4"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft426{font-size:30px;font-family:AAAAAA+OpenSauceOne;color:#a6a6a6;} .ft427{font-size:15px;font-family:CAAAAA+OpenSauceOne;color:#1a1a1a;} .ft428{font-size:15px;font-family:DAAAAA+OpenSauceOne;color:#1a1a1a;} .ft429{font-size:15px;line-height:20px;font-family:DAAAAA+OpenSauceOne;color:#1a1a1a;}--></style><div id="page4-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index004.png" alt="background image"/><p style="position:absolute;top:36px;left:34px;white-space:nowrap" class="ft426"><b>The&#160;Architecture&#160;Sweet&#160;Spot</b></p><p style="position:absolute;top:1135px;left:870px;white-space:nowrap" class="ft420">3</p><p style="position:absolute;top:671px;left:50px;white-space:nowrap" class="ft427"><i><b>Figure&#160;1.&#160;Error–cost&#160;comparison&#160;across&#160;24&#160;architecture&#160;configurations.</b></i></p><p style="position:absolute;top:671px;left:561px;white-space:nowrap" class="ft428"><i>&#160;Each&#160;point&#160;represents&#160;one&#160;architecture</i></p><p style="position:absolute;top:691px;left:50px;white-space:nowrap" class="ft429"><i>evaluated&#160;for&#160;USDJPY&#160;volatility&#160;forecasting.&#160;Cost&#160;reflects&#160;estimated&#160;LLM&#160;inference&#160;cost,&#160;while&#160;error&#160;is&#160;calculated<br/>using&#160;the&#160;quasi-likelihood&#160;loss&#160;measure&#160;described&#160;earlier.&#160;Stars&#160;identify&#160;configurations&#160;that&#160;are&#160;nondominated&#160;(most<br/>optimal)&#160;when&#160;considering&#160;error&#160;and&#160;cost.&#160;The&#160;24&#160;configurations&#160;are&#160;grouped&#160;into&#160;the&#160;A0,&#160;A1,&#160;A2,&#160;and&#160;A3&#160;architecture<br/>families&#160;described&#160;in&#160;Table&#160;1;&#160;the&#160;appendix&#160;provides&#160;details&#160;of&#160;each&#160;variation.</i></p><p style="position:absolute;top:811px;left:50px;white-space:nowrap" class="ft414"><b>Finding&#160;1:&#160;Expect&#160;a&#160;sweet&#160;spot,&#160;not&#160;“more&#160;agents&#160;is&#160;better”</b></p><p style="position:absolute;top:855px;left:50px;white-space:nowrap" class="ft423">The&#160;error–cost&#160;view&#160;in&#160;Figure&#160;1&#160;illustrates&#160;the&#160;type&#160;of&#160;trade-off&#160;product&#160;teams&#160;should&#160;examine:&#160;lower&#160;error&#160;may<br/>require&#160;additional&#160;cost,&#160;but&#160;the&#160;marginal&#160;benefit&#160;can&#160;flatten&#160;or&#160;reverse.&#160;In&#160;the&#160;broader&#160;24-configuration&#160;comparison,<br/>the&#160;lowest-error&#160;candidates&#160;included&#160;A0-GARCH-X,&#160;A1-LLM-Residual,&#160;A1-Probabilistic-LLM,&#160;the&#160;event-only&#160;hybrid,&#160;and<br/>A2-Conditional-Router.&#160;Because&#160;these&#160;configurations&#160;were&#160;evaluated&#160;on&#160;a&#160;very&#160;specific&#160;case&#160;(USDJPY&#160;volatility<br/>forecasting),&#160;the&#160;broader&#160;comparison&#160;should&#160;be&#160;interpreted&#160;as&#160;an&#160;architecture-screening&#160;exercise&#160;rather&#160;than&#160;as&#160;a<br/>single&#160;definitive&#160;ranking.&#160;In&#160;the&#160;full&#160;architecture&#160;comparison,&#160;A1&#160;was&#160;the&#160;strongest&#160;measured&#160;operating&#160;point&#160;and<br/>outperformed&#160;A2&#160;on&#160;error,&#160;capability&#160;score,&#160;and&#160;cost.</p><p style="position:absolute;top:1017px;left:50px;white-space:nowrap" class="ft423">This&#160;is&#160;the&#160;practical&#160;meaning&#160;of&#160;a&#160;sweet&#160;spot.&#160;It&#160;is&#160;not&#160;a&#160;universal&#160;curve&#160;estimated&#160;from&#160;the&#160;data.&#160;It&#160;is&#160;a&#160;set&#160;of<br/>configurations&#160;that&#160;remain&#160;attractive&#160;under&#160;explicit&#160;error,&#160;capability,&#160;and&#160;cost&#160;comparisons.&#160;Product&#160;teams&#160;should<br/>therefore&#160;benchmark&#160;an&#160;architecture&#160;ladder,&#160;measure&#160;marginal&#160;value&#160;per&#160;unit&#160;cost&#160;and&#160;latency,&#160;and&#160;stop&#160;adding<br/>orchestration&#160;when&#160;the&#160;next&#160;stage&#160;does&#160;not&#160;improve&#160;the&#160;target&#160;objective.</p></div><!-- Page 5 --><a name="5"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft530{font-size:14px;line-height:19px;font-family:DAAAAA+TTHoves;color:#000000;}--></style><div id="page5-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index005.png" alt="background image"/><p style="position:absolute;top:36px;left:34px;white-space:nowrap" class="ft526"><b>Domain&#160;Expertise&#160;Determines&#160;Architecture</b></p><p style="position:absolute;top:1135px;left:870px;white-space:nowrap" class="ft520">4</p><p style="position:absolute;top:114px;left:34px;white-space:nowrap" class="ft514"><b>Finding&#160;2:&#160;Domain&#160;expertise&#160;determines&#160;the&#160;architecture</b></p><p style="position:absolute;top:157px;left:34px;white-space:nowrap" class="ft523">The&#160;architecture&#160;worked&#160;because&#160;the&#160;task&#160;was&#160;structured.&#160;The&#160;target&#160;was&#160;realized&#160;volatility,&#160;the&#160;numerical&#160;output<br/>was&#160;positive&#160;and&#160;bounded,&#160;and&#160;the&#160;model&#160;could&#160;act&#160;only&#160;through&#160;a&#160;small&#160;multiplier&#160;grid.&#160;That&#160;made&#160;a&#160;quant&#160;library&#160;the<br/>correct&#160;source&#160;of&#160;arithmetic&#160;and&#160;made&#160;the&#160;language&#160;model’s&#160;role&#160;narrow:&#160;recognize&#160;contextual&#160;conditions&#160;that&#160;might<br/>justify&#160;a&#160;controlled&#160;adjustment.</p><p style="position:absolute;top:259px;left:34px;white-space:nowrap" class="ft523">The&#160;same&#160;design&#160;should&#160;not&#160;be&#160;copied&#160;into&#160;a&#160;task&#160;whose&#160;difficulty&#160;lies&#160;in&#160;evidence&#160;reconciliation.&#160;If&#160;the&#160;business<br/>problem&#160;requires&#160;comparing&#160;contradictory&#160;contracts,&#160;extracting&#160;numeric&#160;exceptions&#160;from&#160;documents,&#160;or<br/>coordinating&#160;tools&#160;with&#160;different&#160;permissions,&#160;a&#160;multi-agent&#160;design&#160;may&#160;be&#160;justified.&#160;But&#160;those&#160;capabilities&#160;must&#160;be<br/>measured&#160;directly.&#160;A&#160;capability&#160;score&#160;inferred&#160;from&#160;model&#160;breadth&#160;or&#160;forecast&#160;loss&#160;is&#160;not&#160;enough.</p><p style="position:absolute;top:403px;left:168px;white-space:nowrap" class="ft516"><b>Task</b></p><p style="position:absolute;top:403px;left:400px;white-space:nowrap" class="ft516"><b>Preferred&#160;Owner</b></p><p style="position:absolute;top:403px;left:696px;white-space:nowrap" class="ft516"><b>Reason</b></p><p style="position:absolute;top:461px;left:74px;white-space:nowrap" class="ft517">Returns,&#160;realized&#160;volatility,&#160;forecasts</p><p style="position:absolute;top:471px;left:380px;white-space:nowrap" class="ft517">Quant/statistical&#160;libraries</p><p style="position:absolute;top:461px;left:645px;white-space:nowrap" class="ft517">Auditable&#160;arithmetic&#160;and</p><p style="position:absolute;top:480px;left:677px;white-space:nowrap" class="ft517">reproducibility</p><p style="position:absolute;top:550px;left:63px;white-space:nowrap" class="ft517">Losses,&#160;confidence&#160;intervals,&#160;backtests</p><p style="position:absolute;top:550px;left:380px;white-space:nowrap" class="ft517">Quant/statistical&#160;libraries</p><p style="position:absolute;top:540px;left:625px;white-space:nowrap" class="ft517">Explicit&#160;definitions&#160;and&#160;testable</p><p style="position:absolute;top:559px;left:692px;white-space:nowrap" class="ft517">inference</p><p style="position:absolute;top:629px;left:78px;white-space:nowrap" class="ft517">Event&#160;classification&#160;and&#160;relevance</p><p style="position:absolute;top:629px;left:379px;white-space:nowrap" class="ft517">Constrained&#160;LLM&#160;or&#160;rules</p><p style="position:absolute;top:619px;left:624px;white-space:nowrap" class="ft517">Context&#160;interpretation&#160;under&#160;an</p><p style="position:absolute;top:638px;left:693px;white-space:nowrap" class="ft517">allow-list</p><p style="position:absolute;top:698px;left:83px;white-space:nowrap" class="ft517">Contradictory-source&#160;resolution</p><p style="position:absolute;top:708px;left:359px;white-space:nowrap" class="ft517">Multi-agent&#160;workflow,&#160;if&#160;needed</p><p style="position:absolute;top:698px;left:614px;white-space:nowrap" class="ft517">Independent&#160;evidence&#160;checks&#160;and</p><p style="position:absolute;top:717px;left:689px;white-space:nowrap" class="ft517">escalation</p><p style="position:absolute;top:794px;left:147px;white-space:nowrap" class="ft517">Final&#160;action</p><p style="position:absolute;top:794px;left:334px;white-space:nowrap" class="ft517">Deterministic&#160;policy&#160;layer</p><p style="position:absolute;top:785px;left:612px;white-space:nowrap" class="ft530">Bounded,&#160;logged,&#160;reversible<br/>decision</p><p style="position:absolute;top:913px;left:41px;white-space:nowrap" class="ft521"><i><b>Table&#160;2:&#160;&#160;Agentic&#160;division&#160;of&#160;labor&#160;-</b></i></p><p style="position:absolute;top:913px;left:290px;white-space:nowrap" class="ft522"><i>&#160;This&#160;table&#160;illustrates&#160;a&#160;division-of-labor&#160;principle&#160;for&#160;agentic&#160;systems.&#160;Traditional</i></p><p style="position:absolute;top:933px;left:41px;white-space:nowrap" class="ft525"><i>(i.e.,&#160;pre-GenAI)&#160;quantitative/machine-learning&#160;libraries&#160;should&#160;handle&#160;calculations&#160;and&#160;statistical&#160;inference&#160;because<br/>these&#160;tasks&#160;require&#160;auditability,&#160;reproducibility,&#160;and&#160;explicit&#160;definitions.&#160;LLMs&#160;or&#160;rules&#160;should&#160;be&#160;used&#160;for&#160;contextual<br/>interpretation,&#160;while&#160;multi-agent&#160;workflows&#160;should&#160;be&#160;reserved&#160;for&#160;genuinely&#160;difficult&#160;evidence-reconciliation&#160;tasks.<br/>Final&#160;actions&#160;should&#160;pass&#160;through&#160;a&#160;bounded,&#160;logged,&#160;and&#160;reversible&#160;policy&#160;layer.</i></p></div><!-- Page 6 --><a name="6"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft631{font-size:15px;font-family:AAAAAA+OpenSauceOne;color:#000000;} .ft632{font-size:18px;font-family:AAAAAA+OpenSauceOne;color:#535353;}--></style><div id="page6-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index006.png" alt="background image"/><p style="position:absolute;top:36px;left:34px;white-space:nowrap" class="ft626"><b>From&#160;Findings&#160;to&#160;Design&#160;Principles</b></p><p style="position:absolute;top:1135px;left:870px;white-space:nowrap" class="ft620">5</p><p style="position:absolute;top:318px;left:45px;white-space:nowrap" class="ft614"><b>Product&#160;and&#160;quantitative&#160;design&#160;guidance</b></p><p style="position:absolute;top:361px;left:45px;white-space:nowrap" class="ft615">The&#160;main&#160;design&#160;lessons&#160;from&#160;this&#160;comparative&#160;study&#160;are:</p><p style="position:absolute;top:402px;left:71px;white-space:nowrap" class="ft631"><b>Begin&#160;with&#160;a&#160;strong&#160;deterministic&#160;baseline.&#160;</b></p><p style="position:absolute;top:402px;left:382px;white-space:nowrap" class="ft615">Calibrate&#160;it&#160;on&#160;a&#160;development&#160;sample&#160;and&#160;make&#160;it&#160;strong&#160;enough&#160;that</p><p style="position:absolute;top:422px;left:71px;white-space:nowrap" class="ft615">any&#160;claimed&#160;LLM&#160;improvement&#160;has&#160;a&#160;credible&#160;benchmark.</p><p style="position:absolute;top:442px;left:71px;white-space:nowrap" class="ft631"><b>Define&#160;the&#160;information&#160;boundary&#160;before&#160;designing&#160;prompts.</b></p><p style="position:absolute;top:442px;left:500px;white-space:nowrap" class="ft615">&#160;A&#160;clear&#160;division&#160;of&#160;labor&#160;between&#160;quantitative</p><p style="position:absolute;top:463px;left:71px;white-space:nowrap" class="ft615">systems&#160;and&#160;LLMs&#160;is&#160;more&#160;important&#160;than&#160;prompt&#160;sophistication&#160;alone.</p><p style="position:absolute;top:483px;left:71px;white-space:nowrap" class="ft631"><b>Use&#160;the&#160;smallest&#160;architecture&#160;that&#160;can&#160;complete&#160;the&#160;task.</b></p><p style="position:absolute;top:483px;left:488px;white-space:nowrap" class="ft615">&#160;Add&#160;agents&#160;only&#160;when&#160;they&#160;perform&#160;a&#160;distinct</p><p style="position:absolute;top:503px;left:71px;white-space:nowrap" class="ft615">operation&#160;that&#160;can&#160;be&#160;measured&#160;independently.</p><p style="position:absolute;top:523px;left:71px;white-space:nowrap" class="ft631"><b>Keep&#160;the&#160;action&#160;space&#160;bounded&#160;when&#160;the&#160;business&#160;process&#160;permits&#160;it.&#160;</b></p><p style="position:absolute;top:523px;left:586px;white-space:nowrap" class="ft615">A&#160;constrained&#160;multiplier&#160;or&#160;policy&#160;choice</p><p style="position:absolute;top:544px;left:71px;white-space:nowrap" class="ft615">is&#160;easier&#160;to&#160;validate&#160;than&#160;unconstrained&#160;numerical&#160;or&#160;textual&#160;generation.</p><p style="position:absolute;top:564px;left:71px;white-space:nowrap" class="ft631"><b>Create&#160;a&#160;separate&#160;capability&#160;suite.</b></p><p style="position:absolute;top:564px;left:319px;white-space:nowrap" class="ft615">&#160;Test&#160;extraction,&#160;conflict&#160;detection,&#160;citation&#160;accuracy,&#160;tool&#160;compliance,&#160;and</p><p style="position:absolute;top:584px;left:71px;white-space:nowrap" class="ft623">temporal&#160;reasoning&#160;using&#160;known&#160;answers&#160;rather&#160;than&#160;inferring&#160;capability&#160;from&#160;model&#160;size&#160;or&#160;forecast<br/>performance.</p><p style="position:absolute;top:663px;left:45px;white-space:nowrap" class="ft614"><b>Conclusion</b></p><p style="position:absolute;top:711px;left:45px;white-space:nowrap" class="ft615">The&#160;central&#160;lesson&#160;from&#160;this&#160;study&#160;is&#160;</p><p style="position:absolute;top:711px;left:304px;white-space:nowrap" class="ft631"><b>architectural&#160;discipline</b></p><p style="position:absolute;top:711px;left:471px;white-space:nowrap" class="ft615">.&#160;Agentic&#160;systems&#160;can&#160;have&#160;a&#160;sweet&#160;spot,&#160;but&#160;that&#160;sweet</p><p style="position:absolute;top:732px;left:45px;white-space:nowrap" class="ft623">spot&#160;is&#160;created&#160;by&#160;matching&#160;the&#160;architecture&#160;to&#160;the&#160;domain&#160;rather&#160;than&#160;by&#160;maximizing&#160;the&#160;number&#160;of&#160;agents&#160;or&#160;the<br/>amount&#160;of&#160;LLM&#160;reasoning.&#160;Traditional&#160;quantitative&#160;libraries&#160;should&#160;own&#160;arithmetic&#160;and&#160;statistical&#160;calculations&#160;because<br/>those&#160;tasks&#160;are&#160;auditable,&#160;reproducible,&#160;and&#160;robust.&#160;LLMs&#160;should&#160;be&#160;reserved&#160;for&#160;contextual&#160;tasks&#160;that&#160;genuinely<br/>require&#160;interpretation&#160;or&#160;reasoning.&#160;The&#160;additional&#160;context&#160;should&#160;then&#160;be&#160;measured&#160;against&#160;the&#160;business&#160;objective&#160;to<br/>determine&#160;whether&#160;it&#160;justifies&#160;the&#160;added&#160;cost&#160;and&#160;operational&#160;complexity.</p><p style="position:absolute;top:131px;left:45px;white-space:nowrap" class="ft614"><b>Finding&#160;3:&#160;Calculation&#160;versus&#160;context</b></p><p style="position:absolute;top:175px;left:45px;white-space:nowrap" class="ft623">A&#160;robust&#160;agentic&#160;system&#160;should&#160;divide&#160;work&#160;according&#160;to&#160;the&#160;strengths&#160;of&#160;each&#160;component.&#160;Deterministic&#160;libraries<br/>should&#160;own&#160;calculations&#160;that&#160;have&#160;explicit&#160;definitions,&#160;stable&#160;tests,&#160;and&#160;known&#160;numerical&#160;behavior.&#160;The&#160;language<br/>model&#160;should&#160;handle&#160;contextual&#160;interpretation&#160;only&#160;where&#160;it&#160;adds&#160;information&#160;that&#160;the&#160;numerical&#160;system&#160;cannot<br/>obtain&#160;from&#160;structured&#160;inputs.&#160;In&#160;this&#160;study,&#160;adding&#160;more&#160;reasoning&#160;stages&#160;to&#160;the&#160;LLM&#160;component&#160;did&#160;not&#160;improve<br/>measured&#160;forecast&#160;error&#160;enough&#160;to&#160;justify&#160;the&#160;additional&#160;cost&#160;and&#160;complexity.</p><p style="position:absolute;top:928px;left:45px;white-space:nowrap" class="ft632"><b>Data&#160;and&#160;Production&#160;Context</b></p><p style="position:absolute;top:976px;left:45px;white-space:nowrap" class="ft623">Real-time&#160;market&#160;updates&#160;used&#160;by&#160;the&#160;production-oriented&#160;workflow&#160;were&#160;sourced&#160;through&#160;EMAlpha’s&#160;CoTrader<br/>agent.&#160;The&#160;historical&#160;evaluation&#160;was&#160;replayed&#160;on&#160;a&#160;frozen,&#160;point-in-time&#160;dataset&#160;so&#160;that&#160;the&#160;results&#160;could&#160;be&#160;audited<br/>and&#160;reproduced.&#160;Quantitative&#160;libraries&#160;remained&#160;responsible&#160;for&#160;the&#160;numerical&#160;calculations,&#160;forecast&#160;generation,&#160;and<br/>error&#160;evaluation.</p></div><!-- Page 7 --><a name="7"></a><style type="text/css"><!-- p {margin: 0; padding: 0;}--></style><div id="page7-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index007.png" alt="background image"/><p style="position:absolute;top:28px;left:30px;white-space:nowrap" class="ft726"><b>Appendix:&#160;Description&#160;of&#160;Architectures</b></p><p style="position:absolute;top:1135px;left:870px;white-space:nowrap" class="ft720">6</p><p style="position:absolute;top:1025px;left:46px;white-space:nowrap" class="ft727"><i><b>Table&#160;A.&#160;Summary&#160;of&#160;the&#160;24&#160;architecture&#160;configurations.</b></i></p><p style="position:absolute;top:1025px;left:451px;white-space:nowrap" class="ft728"><i>&#160;Error&#160;is&#160;measured&#160;using&#160;the&#160;QLIKE&#160;quasi-likelihood&#160;loss,</i></p><p style="position:absolute;top:1045px;left:46px;white-space:nowrap" class="ft729"><i>which&#160;compares&#160;each&#160;predicted&#160;volatility&#160;with&#160;the&#160;subsequently&#160;realized&#160;volatility;&#160;lower&#160;values&#160;indicate&#160;better<br/>forecasts.&#160;Capability&#160;is&#160;an&#160;exploratory&#160;score&#160;between&#160;0&#160;and&#160;1&#160;based&#160;on&#160;observable&#160;features&#160;such&#160;as&#160;contextual<br/>coverage,&#160;event&#160;handling,&#160;specialist&#160;diversity,&#160;synthesis&#160;quality,&#160;critic&#160;behavior,&#160;and&#160;compliance&#160;with&#160;output<br/>constraints.&#160;Estimated&#160;cost&#160;represents&#160;the&#160;LLM&#160;inference&#160;cost&#160;per&#160;forecast&#160;case,&#160;calculated&#160;from&#160;logged&#160;input<br/>and&#160;output&#160;tokens&#160;and&#160;the&#160;applicable&#160;model&#160;prices;&#160;it&#160;excludes&#160;engineering,&#160;infrastructure,&#160;storage,&#160;and&#160;local<br/>quantitative-computing&#160;costs</i></p></div><!-- Page 8 --><a name="8"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft833{font-size:15px;font-family:BAAAAA+OpenSauceOne;color:#1a1a1a;} .ft834{font-size:15px;line-height:20px;font-family:BAAAAA+OpenSauceOne;color:#1a1a1a;}--></style><div id="page8-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index008.png" alt="background image"/><p style="position:absolute;top:36px;left:34px;white-space:nowrap" class="ft826"><b>Appendix:&#160;Detailed&#160;Architecture&#160;Descriptions</b></p><p style="position:absolute;top:1135px;left:871px;white-space:nowrap" class="ft820">7</p><p style="position:absolute;top:124px;left:32px;white-space:nowrap" class="ft831"><b>Detailed&#160;Descriptions&#160;of&#160;the&#160;Agentic&#160;Architectures</b></p><p style="position:absolute;top:164px;left:32px;white-space:nowrap" class="ft834">Table&#160;A&#160;(previous&#160;page)&#160;brings&#160;together&#160;24&#160;architecture&#160;configurations&#160;spanning&#160;four&#160;levels&#160;of&#160;computational&#160;and<br/>agentic&#160;complexity.&#160;The&#160;purpose&#160;of&#160;the&#160;comparative&#160;study&#160;in&#160;this&#160;white&#160;paper&#160;is&#160;not&#160;simply&#160;to&#160;compare&#160;different<br/>model&#160;names,&#160;but&#160;to&#160;examine&#160;how&#160;the&#160;division&#160;of&#160;labor&#160;between&#160;quantitative&#160;models,&#160;single-call&#160;LLMs,&#160;and&#160;multi-<br/>agent&#160;workflows&#160;affects&#160;forecast&#160;error,&#160;capability,&#160;and&#160;estimated&#160;inference&#160;cost.</p><p style="position:absolute;top:266px;left:32px;white-space:nowrap" class="ft834">The&#160;A0&#160;group&#160;contains&#160;nine&#160;deterministic&#160;quantitative&#160;configurations.&#160;These&#160;include&#160;conventional&#160;volatility&#160;models<br/>such&#160;as&#160;EWMA,&#160;GARCH,&#160;HAR,&#160;and&#160;GARCH-X,&#160;as&#160;well&#160;as&#160;richer&#160;quantitative&#160;variants&#160;such&#160;as&#160;regime-based&#160;GARCH,<br/>multivariate&#160;DCC-style&#160;modeling,&#160;and&#160;a&#160;realized-volatility&#160;ensemble.&#160;The&#160;A0&#160;family&#160;also&#160;includes&#160;two&#160;hybrids&#160;in&#160;which<br/>an&#160;LLM&#160;is&#160;added&#160;to&#160;the&#160;GARCH-X&#160;baseline&#160;for&#160;either&#160;event-only&#160;or&#160;market-only&#160;interpretation.&#160;These&#160;configurations<br/>establish&#160;the&#160;numerical&#160;and&#160;calibration&#160;benchmarks&#160;against&#160;which&#160;the&#160;LLM-based&#160;architectures&#160;can&#160;be&#160;evaluated.</p><p style="position:absolute;top:387px;left:32px;white-space:nowrap" class="ft834">The&#160;A1&#160;group&#160;contains&#160;seven&#160;light-LLM&#160;configurations&#160;that&#160;make&#160;targeted&#160;adjustments&#160;to&#160;quantitative&#160;forecasts.&#160;The<br/>benchmark&#160;versions&#160;use&#160;different&#160;model&#160;sizes&#160;-&#160;Nano,&#160;Mini,&#160;GPT-5,&#160;and&#160;GPT-5.5&#160;-&#160;to&#160;test&#160;whether&#160;a&#160;larger&#160;or&#160;more<br/>expensive&#160;model&#160;produces&#160;enough&#160;improvement&#160;to&#160;justify&#160;its&#160;cost.&#160;The&#160;additional&#160;pilot&#160;configurations&#160;assign&#160;the&#160;LLM<br/>more&#160;specific&#160;roles:&#160;classifying&#160;event&#160;relevance,&#160;modeling&#160;a&#160;residual&#160;adjustment,&#160;or&#160;producing&#160;a&#160;probabilistic<br/>forecast.&#160;These&#160;are&#160;deliberately&#160;narrow,&#160;single-call&#160;designs.&#160;They&#160;use&#160;the&#160;quantitative&#160;system&#160;as&#160;the&#160;primary<br/>numerical&#160;engine&#160;and&#160;ask&#160;the&#160;LLM&#160;to&#160;perform&#160;one&#160;clearly&#160;bounded&#160;contextual&#160;or&#160;calibration&#160;task.</p><p style="position:absolute;top:529px;left:32px;white-space:nowrap" class="ft834">The&#160;A2&#160;group&#160;distributes&#160;the&#160;task&#160;across&#160;multiple&#160;LLM-based&#160;decision&#160;stages.&#160;It&#160;contains&#160;six&#160;multi-agent<br/>configurations.&#160;The&#160;original&#160;A2&#160;design&#160;used&#160;several&#160;specialists&#160;followed&#160;by&#160;synthesis&#160;and&#160;critique.&#160;The&#160;event-priority<br/>versions&#160;introduced&#160;a&#160;protocol&#160;that&#160;explicitly&#160;instructed&#160;the&#160;system&#160;to&#160;give&#160;precedence&#160;to&#160;scheduled-event<br/>information&#160;and&#160;included&#160;either&#160;a&#160;critic&#160;or&#160;no&#160;critic.&#160;The&#160;other&#160;variants&#160;explored&#160;alternative&#160;coordination&#160;strategies:<br/>competitive&#160;specialists,&#160;independent&#160;synthesis,&#160;and&#160;a&#160;conditional&#160;router&#160;that&#160;selects&#160;a&#160;workflow&#160;based&#160;on&#160;the&#160;case.<br/>These&#160;configurations&#160;test&#160;whether&#160;additional&#160;agents&#160;and&#160;reasoning&#160;stages&#160;create&#160;enough&#160;incremental&#160;value&#160;to&#160;justify<br/>their&#160;higher&#160;cost&#160;and&#160;operational&#160;complexity.</p><p style="position:absolute;top:691px;left:32px;white-space:nowrap" class="ft834">The&#160;A3&#160;group&#160;gives&#160;the&#160;LLM&#160;greater&#160;responsibility&#160;for&#160;generating&#160;the&#160;forecast&#160;directly,&#160;rather&#160;than&#160;using&#160;it&#160;primarily&#160;to<br/>adjust&#160;a&#160;quantitative&#160;model.&#160;It&#160;contains&#160;two&#160;direct-forecast&#160;configurations.&#160;These&#160;designs&#160;reduce&#160;or&#160;remove&#160;the&#160;role<br/>of&#160;the&#160;quantitative&#160;volatility&#160;anchor&#160;and&#160;ask&#160;the&#160;LLM&#160;to&#160;produce&#160;the&#160;forecast&#160;more&#160;directly.&#160;The&#160;benchmark&#160;A3&#160;uses<br/>GPT-5,&#160;while&#160;the&#160;pilot&#160;A3&#160;adds&#160;explicit&#160;output&#160;constraints.&#160;This&#160;group&#160;tests&#160;the&#160;opposite&#160;design&#160;philosophy&#160;from&#160;A1<br/>and&#160;A2:&#160;rather&#160;than&#160;using&#160;the&#160;LLM&#160;to&#160;adjust&#160;a&#160;trusted&#160;numerical&#160;forecast,&#160;it&#160;gives&#160;the&#160;LLM&#160;greater&#160;responsibility&#160;for<br/>generating&#160;the&#160;forecast&#160;itself.</p><p style="position:absolute;top:833px;left:32px;white-space:nowrap" class="ft834">The&#160;Capability&#160;score&#160;is&#160;an&#160;exploratory&#160;measure&#160;of&#160;how&#160;much&#160;contextual&#160;and&#160;orchestration&#160;functionality&#160;an<br/>architecture&#160;can&#160;perform&#160;beyond&#160;numerical&#160;forecasting.&#160;It&#160;combines&#160;observable&#160;components&#160;such&#160;as&#160;event<br/>coverage,&#160;contextual&#160;relevance,&#160;specialist&#160;diversity,&#160;synthesis&#160;quality,&#160;critic&#160;behavior,&#160;and&#160;compliance&#160;with&#160;output<br/>constraints.&#160;These&#160;components&#160;are&#160;normalized&#160;to&#160;a&#160;common&#160;scale&#160;and&#160;combined&#160;into&#160;a&#160;weighted&#160;score&#160;between&#160;0<br/>and&#160;1,&#160;where&#160;higher&#160;values&#160;indicate&#160;broader&#160;measured&#160;capability.&#160;The&#160;score&#160;should&#160;be&#160;interpreted&#160;directionally&#160;rather<br/>than&#160;as&#160;a&#160;universal&#160;measure&#160;of&#160;intelligence,&#160;because&#160;not&#160;every&#160;architecture&#160;was&#160;evaluated&#160;on&#160;an&#160;identical&#160;standalone<br/>capability&#160;test&#160;suite.</p></div><!-- Page 9 --><a name="9"></a><style type="text/css"><!-- p {margin: 0; padding: 0;} .ft935{font-size:21px;font-family:AAAAAA+TTHoves;color:#535353;} .ft936{font-size:21px;font-family:BAAAAA+TTHoves;color:#535353;} .ft937{font-size:22px;font-family:BAAAAA+TTHoves;color:#535353;} .ft938{font-size:22px;font-family:AAAAAA+TTHoves;color:#535353;} .ft939{font-size:30px;font-family:CAAAAA+OpenSauceOne;color:#a6a6a6;} .ft940{font-size:22px;line-height:37px;font-family:BAAAAA+TTHoves;color:#535353;}--></style><div id="page9-div" style="position:relative;width:918px;height:1187px;"><img width="918" height="1187" src="index009.png" alt="background image"/><p style="position:absolute;top:613px;left:343px;white-space:nowrap" class="ft935"><b>Thank&#160;you&#160;for&#160;reading.</b></p><p style="position:absolute;top:667px;left:343px;white-space:nowrap" class="ft936">We&#160;welcome&#160;questions&#160;and&#160;conversations&#160;about</p><p style="position:absolute;top:693px;left:343px;white-space:nowrap" class="ft936">designing&#160;cost-effective,&#160;reliable,&#160;and&#160;domain-</p><p style="position:absolute;top:720px;left:343px;white-space:nowrap" class="ft936">appropriate&#160;agentic&#160;systems&#160;for&#160;enterprise&#160;applications.</p><p style="position:absolute;top:747px;left:343px;white-space:nowrap" class="ft936">EMAlpha&#160;creates&#160;industry-focused&#160;agents&#160;that</p><p style="position:absolute;top:773px;left:343px;white-space:nowrap" class="ft936">combine&#160;real-time&#160;contextual&#160;intelligence&#160;with&#160;trusted</p><p style="position:absolute;top:800px;left:343px;white-space:nowrap" class="ft936">quantitative&#160;and&#160;operational&#160;systems.&#160;Our&#160;CoTrader</p><p style="position:absolute;top:827px;left:343px;white-space:nowrap" class="ft936">agent&#160;is&#160;an&#160;example&#160;of&#160;this&#160;approach&#160;for&#160;financial-</p><p style="position:absolute;top:854px;left:343px;white-space:nowrap" class="ft936">market&#160;workflows.&#160;Please&#160;contact&#160;us&#160;to&#160;discuss&#160;the</p><p style="position:absolute;top:880px;left:343px;white-space:nowrap" class="ft936">findings,&#160;the&#160;experiment,&#160;or&#160;potential&#160;applications&#160;to</p><p style="position:absolute;top:907px;left:343px;white-space:nowrap" class="ft936">your&#160;organization.</p><p style="position:absolute;top:977px;left:394px;white-space:nowrap" class="ft940">10&#160;Glenlake&#160;Parkway,&#160;Suite&#160;130,&#160;<br/>Atlanta,&#160;GA&#160;30328,&#160;USA</p><p style="position:absolute;top:1052px;left:394px;white-space:nowrap" class="ft938"><b>[email protected]</b></p><p style="position:absolute;top:1090px;left:394px;white-space:nowrap" class="ft937">https://emalpha.com</p><p style="position:absolute;top:269px;left:61px;white-space:nowrap" class="ft939"><b>·&#160;Agentic&#160;architecture&#160;design&#160;</b></p><p style="position:absolute;top:328px;left:61px;white-space:nowrap" class="ft939"><b>·&#160;Quantitative&#160;systems&#160;</b></p><p style="position:absolute;top:386px;left:61px;white-space:nowrap" class="ft939"><b>·&#160;LLM&#160;governance&#160;</b></p><p style="position:absolute;top:445px;left:61px;white-space:nowrap" class="ft939"><b>·&#160;Cost–quality-capability&#160;trade-offs</b></p></div><hr/><a name="outline"></a><h1>Document Outline</h1><ul><li><a href="index.html#1">Finding the Agentic Sweet Spot</a><ul><li><a href="index.html#2">Enterprise Design Constraints</a></li></ul></li><li><a href="index.html#2">Executive Summary</a><ul><li><a href="index.html#2">Choosing the Right Agentic Architecture for Enterprise Tasks</a><ul><li><a href="index.html#2">Enterprise agentic systems rarely improve monotonically as more agents, tools, and reasoning stages are added. In this white paper, we present a comparative study of agentic architectures of varying complexity for predicting currency volatility. In this experiment, a simple hybrid combining a traditional quantitative model with a low-cost LLM significantly improved the results, while full orchestration of LLM-based agents was substantially more expensive and less accurate. The result is not a universal ranking of models. It is a design lesson: the most efficient agentic architecture depends on the problem, the information available at decision time, and the appropriate division of tasks between traditional models and LLMs.</a><ul><li><a href="index.html#2">The study led to the following lessons:</a></li><li><a href="index.html#2">An agentic sweet spot is plausible: The observed configurations formed a practical cost–accuracy trade-off. The light hybrid achieved better accuracy at an estimated cost of $0.015 per case. The orchestrated hybrid delivered lower accuracy at a higher estimated cost of $0.093 per case. In this experiment, additional orchestration did not produce better measured performance.</a></li><li><a href="index.html#2">Domain structure determines architecture: The system forecast one-hour-ahead USDJPY realized volatility around scheduled macroeconomic events. Quantitative finance provided natural numerical-model anchors, a known event calendar, and a bounded action space. A different problem - such as multi-document reconciliation or policy interpretation - could justify a richer agentic design.</a></li><li><a href="index.html#2">Separate calculation from context: Quant libraries should calculate returns, realized volatility, GARCH forecasts, losses, and statistical tests. LLMs should interpret contextual evidence, identify relevant regimes, and select among pre-approved actions. LLMs should not replace arithmetic that is already auditable and robust.</a></li></ul></li></ul></li><li><a href="index.html#2">Budget</a><ul><li><a href="index.html#2">Operate within the allocated budget.</a></li></ul></li><li><a href="index.html#2">Quality</a><ul><li><a href="index.html#2">Meet defined quality and accuracy benchmarks.</a></li></ul></li><li><a href="index.html#2">Compliance</a><ul><li><a href="index.html#2">Adhere to governance and compliance guardrails.</a></li></ul></li></ul></li><li><a href="index.html#3">Matching Agentic Architecture to Enterprise Requirements</a><ul><li><a href="index.html#3">Lessons from a quant–LLM experiment in architecture, cost, and quality</a></li><li><a href="index.html#3">Finding the Right Architecture</a><ul><li><a href="index.html#3">Architecture Family</a></li><li><a href="index.html#3">Role of LLM</a></li><li><a href="index.html#3">Error</a></li><li><a href="index.html#3">Estimated cost/case</a></li></ul></li></ul></li><li><a href="index.html#4">The Architecture Sweet Spot</a><ul><li><a href="index.html#4">Finding 1: Expect a sweet spot, not “more agents is better”</a></li></ul></li><li><a href="index.html#5">Domain Expertise Determines Architecture</a><ul><li><a href="index.html#5">Finding 2: Domain expertise determines the architecture</a><ul><li><a href="index.html#5">Task</a></li><li><a href="index.html#5">Preferred Owner</a></li><li><a href="index.html#5">Reason</a></li></ul></li></ul></li><li><a href="index.html#6">From Findings to Design Principles</a><ul><li><a href="index.html#6">Finding 3: Calculation versus context</a></li><li><a href="index.html#6">Product and quantitative design guidance</a></li><li><a href="index.html#6">Conclusion</a></li><li><a href="index.html#6">Data and Production Context</a></li></ul></li><li><a href="index.html#7">Appendix: Description of Architectures</a></li><li><a href="index.html#8">Appendix: Detailed Architecture Descriptions</a><ul><li><a href="index.html#8">Detailed Descriptions of the Agentic Architectures</a></li></ul></li><li><a href="index.html#9">· Agentic architecture design</a></li><li><a href="index.html#9">· Quantitative systems</a></li><li><a href="index.html#9">· LLM governance</a></li><li><a href="index.html#9">· Cost–quality-capability trade-offs</a><ul><li><a href="index.html#9">Thank you for reading.</a></li><li><a href="index.html#9">[email protected]</a></li></ul></li></ul></body></html>

Related research

More published perspectives.

A Tale of Two Stock Markets
ai-llmJune 24, 2026

A Tale of Two Stock Markets

Bloomberg and Reuters this morning both cited South Korean local media coverage as the trigger behind today’s sharp KOSPI sell-off. But EMAlpha’s MacroMonitor AI Agent had been tracking this narrative

ai-llmemerging-markets
emerging-markets · markets-and-macroView insight