AI-Powered Multi-Agent Decision Support System for Strategic Business Decisions

coinbase · Finance · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise strategic decision support

Coinbase needed AI support for high-stakes RAPID decisions that could expose unseen risks, challenge assumptions, retrieve organizational context, and keep human deciders accountable rather than replacing them.

Constraints:

Architecture

Coinbase's RAPID-D assistant is a modular multi-agent system. A Single Shot Recommender reads the RAPID document and forms a baseline recommendation. A Contextual Recommender generates questions and uses enterprise search to retrieve broader internal context. A Contrarian Agent builds the strongest case against the initial recommendation, probing assumptions and risks. A Debate and Decide Agent synthesizes those views into a final recommendation for the human Decider. The implementation adds asynchronous processing for complex decisions, Claude 3.7 Sonnet for the model layer, human-review benchmarking against real RAPID decisions, and feedback capture from user and stakeholder comments.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  rapid-document["rapid-document"] --> analyst-agent["analyst-agent"]
  rapid-document --> context-retrieval["context-retrieval"]
  context-retrieval --> enterprise-knowledge-base["enterprise-knowledge-base"]
  context-retrieval --> seeker-agent["seeker-agent"]
  analyst-agent --> contrarian-agent["contrarian-agent"]
  analyst-agent --> synthesizer-agent["synthesizer-agent"]
  seeker-agent --> synthesizer-agent
  contrarian-agent --> synthesizer-agent
  synthesizer-agent --> human-decider["human-decider"]
  synthesizer-agent --> async-processing["async-processing"]
  human-decider --> feedback-loop["feedback-loop"]
  feedback-loop --> synthesizer-agent
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
rapid-document Structured decision document containing the RAPID framework inputs. analyst-agent, context-retrieval
analyst-agent Single Shot Recommender that produces a baseline recommendation from the RAPID document. rapid-document, synthesizer-agent
context-retrieval Question generator and enterprise search tool for retrieving decision context. rapid-document, enterprise-knowledge-base, seeker-agent
enterprise-knowledge-base Coinbase internal knowledge sources used to answer context questions. context-retrieval
seeker-agent Contextual Recommender that synthesizes broader organizational context. context-retrieval, synthesizer-agent
contrarian-agent Devil's Advocate that argues against the initial recommendation. analyst-agent, synthesizer-agent
synthesizer-agent Debate and Decide Agent that combines analyst, seeker, and contrarian views. analyst-agent, seeker-agent, contrarian-agent, human-decider
human-decider Accountable RAPID Decider who receives the final recommendation and retains oversight. synthesizer-agent, feedback-loop
feedback-loop Captures comments and corrections from users or stakeholders for subsequent recommendations. human-decider, synthesizer-agent
async-processing Keeps users informed when complex decision analysis requires more time. synthesizer-agent, human-decider

Data flow

A RAPID decision document is given to the assistant. The Analyst agent reviews the document and creates a baseline recommendation from the facts in the document. The Seeker generates targeted questions and retrieves answers from internal knowledge sources. The Devil's Advocate challenges the recommendation by identifying assumptions, weaknesses, risks, and unintended consequences. The Synthesizer evaluates all arguments and sends a reasoned recommendation with tradeoffs to the human Decider. Feedback and corrections are captured against the original recommendation for subsequent optimization.

4 supporting quotes
“This agent first performs a thorough, impartial review of the primary RAPID document”article, stated
“It then leverages our enterprise search tool to find answers across all internal knowledge sources.”article, stated
“It deliberately probes for weaknesses, unstated assumptions, potential risks, and unintended consequences.”article, stated
“It then produces a comprehensive final recommendation for the human Decider, complete with a detailed explanation of its reasoning and the trade-offs it considered”article, stated

Why this architecture

The architecture mirrors an expert advisory group because a single model answer would hide bias and reasoning. Splitting roles makes the system read the facts, gather external enterprise context, argue against itself, and then synthesize. The retrieval tool broadens evidence beyond the RAPID document. The async path acknowledges that rigorous analysis can take time. Human review against real decisions and stakeholder corrections keep recommendations calibrated to Coinbase's decision process.

4 supporting quotes
“RAPID-D isn’t a monolithic AI that provides a single, black-box answer.”article, stated
“It's a modular, multi-agent system where specialized agents collaborate, debate, and synthesize information, mirroring a team of expert advisors.”article, stated
“This process helps ensure recommendations are comprehensive, less biased, and tailored to the specific context of each strategic decision”article, stated
“We also implemented an asynchronous architecture, ensuring that for complex decisions requiring more time, users are kept informed and receive a well-reasoned output without unnecessary delays.”article, stated

Models used: Claude 3.7 Sonnet · large language models

Scale

Alternatives considered (roads not taken)

A monolithic AI answer

Rejected because: Coinbase wanted collaborative, debatable, synthesized reasoning rather than a single black-box recommendation.

“RAPID-D isn’t a monolithic AI that provides a single, black-box answer.”article, stated

The first single-agent prototype

Rejected because: Manual feedback showed deciders needed more interaction, explainability, and multi-agent debate.

“We started with a single agent that analyzed a RAPID document and presented its reasoning. Early feedback was gathered manually”article, stated
“The current version is a far more dynamic and interactive experience. It incorporates the multi-agent debate and provides deeper explainability for the AI’s conclusions.”article, stated

Immediate synchronous output for every decision

Rejected because: Complex decisions can need more time, so users need asynchronous status and reasoned later output.

“for complex decisions requiring more time, users are kept informed and receive a well-reasoned output without unnecessary delays.”article, stated

Patterns

multi-agent-orchestration (orchestration)

RAPID-D coordinates specialized agents for analysis, contextual search, contrarian critique, and synthesis.

“It's a modular, multi-agent system where specialized agents collaborate, debate, and synthesize information, mirroring a team of expert advisors.”article, stated

rag-retrieval (retrieval)

The Seeker uses enterprise search to retrieve relevant internal knowledge for generated critical questions.

“For every question, it retrieves relevant information by searching Coinbase’s enterprise knowledge base.”article, stated

human-in-the-loop-gate (governance)

The output is a recommendation for the human RAPID Decider, and stakeholder corrections feed the process.

“It then produces a comprehensive final recommendation for the human Decider”article, stated
“Comments or corrections — whether provided by the user during an active session or later by any stakeholder in the RAPID document — are captured and analyzed”article, stated

eval-driven-development (evaluation)

The team benchmarked assistant recommendations against real RAPID decider outcomes and used leadership feedback to reshape the design.

“We measured RAPID‑D Assistant’s accuracy through a human review process, comparing each of its final recommendations against the real decisions documented by Coinbase’s RAPID Deciders.”article, stated

Tradeoffs

Use multi-agent debate.

Chose specialized agents that collaborate, debate, and synthesize over a monolithic AI recommendation.

Rationale: The system needed transparent, less biased, reproducible decision analysis.

“This structured, multi-step pipeline ensures a rigorous examination of the facts and trade-offs.”article, stated
“More importantly, it makes the entire decision-making process transparent, consistent, and reproducible.”article, stated

Keep human oversight.

Chose AI-driven diligence engine for a human Decider over automated final decisions.

Rationale: The source explicitly frames RAPID-D as augmentation, not replacement.

“It is not designed to replace human oversight but to augment it with a scalable, AI-driven diligence engine.”article, stated

Choose Claude 3.7 Sonnet.

Chose Claude 3.7 Sonnet over other benchmarked leading models.

Rationale: It balanced quality, stability, and reliability.

“Claude 3.7 Sonnet was ultimately chosen for its strong balance of quality, stability, and reliability.”article, stated

Gotchas (what bit them)

Decision-support AI can become a bias amplifier if it only supports the initial recommendation.

Root cause: A single recommendation path may miss weaknesses, assumptions, and unintended consequences.

Resolution: Introduce an explicit Contrarian Agent that builds the strongest case against the initial recommendation.

“This agent’s sole purpose is to build the strongest possible case against the initial recommendation.”article, stated
“It deliberately probes for weaknesses, unstated assumptions, potential risks, and unintended consequences.”article, stated

Complex decision analysis can exceed a normal interactive response window.

Root cause: Retrieval, debate, and synthesis for strategic decisions require more work than a direct answer.

Resolution: Use asynchronous architecture to keep users informed and return a reasoned output without unnecessary delay.

“We also implemented an asynchronous architecture, ensuring that for complex decisions requiring more time, users are kept informed and receive a well-reasoned output without unnecessary delays.”article, stated

Evaluation

Coinbase measured RAPID-D accuracy through human review against real RAPID decisions, benchmarked leading models, and uses user or stakeholder feedback to optimize subsequent recommendations.

Metrics: human review comparison to real RAPID decisions · benchmark scores across leading models · quality, stability, and reliability · comments and corrections captured during or after the session

“We measured RAPID‑D Assistant’s accuracy through a human review process, comparing each of its final recommendations against the real decisions documented by Coinbase’s RAPID Deciders.”article, stated
“Comments or corrections — whether provided by the user during an active session or later by any stakeholder in the RAPID document — are captured and analyzed”article, stated

Outcomes

The source does not provide adoption counts; the relevant quantified details are a four-agent workflow and use of Claude 3.7 Sonnet.

“More importantly, it makes the entire decision-making process transparent, consistent, and reproducible.”article, stated
“It is not designed to replace human oversight but to augment it with a scalable, AI-driven diligence engine.”article, stated