AI-Powered Multi-Agent Decision Support System for Enterprise Strategic Planning

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 wanted to augment its RAPID decision-making framework with AI that could analyze decision documents, surface missing context and risks, reduce bias, and provide transparent recommendations while leaving final authority with human deciders.

Constraints:

• strategic decisions require accountable human deciders
• recommendations need to be auditable, explainable, and tied to the RAPID document
• bias mitigation requires challenging assumptions rather than only agreeing with the initial recommendation
• enterprise knowledge must be retrieved from internal sources
• complex decisions may take time, so users need asynchronous status and well-reasoned output
• model choice must balance quality, stability, and reliability

Architecture

Coinbase built RAPID-D as a modular multi-agent decision-support assistant. A Single Shot Recommender reviews the primary RAPID document and produces a baseline recommendation. A Contextual Recommender generates critical questions and searches Coinbase's enterprise knowledge base for broader context. A Contrarian Agent intentionally argues against the initial recommendation, probing weaknesses and risks. A Debate and Decide Agent synthesizes the analyst, contextual, and contrarian views into a final recommendation for the human Decider. The system uses document-context retrieval, asynchronous architecture for long-running decisions, human review against real RAPID decisions, Claude 3.7 Sonnet, and feedback capture from users or stakeholders.

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

flowchart LR
  rapid-document["rapid-document"] --> analyst-agent["analyst-agent"]
  rapid-document --> document-context-retrieval["document-context-retrieval"]
  document-context-retrieval --> enterprise-knowledge-base["enterprise-knowledge-base"]
  document-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-status-layer["async-status-layer"]
  human-decider --> feedback-loop["feedback-loop"]
  feedback-loop --> synthesizer-agent

Components

Data flow

A RAPID decision document enters the assistant. The Analyst reviews the document and creates a baseline recommendation. The Seeker generates critical questions and retrieves answers from enterprise knowledge sources. The Devil's Advocate challenges the initial recommendation by looking for weaknesses, assumptions, risks, and unintended consequences. The Synthesizer evaluates all arguments and produces a recommendation with reasoning and tradeoffs for the human Decider. Human review compares recommendations to real RAPID decisions, and comments or corrections are captured to optimize subsequent recommendations.

Why this architecture

Coinbase rejected a monolithic assistant because strategic decisions need multiple perspectives: factual reading, contextual retrieval, adversarial challenge, and synthesis. The agent roles mirror a team of advisors and explicitly combat bias by making one agent argue against the initial recommendation. Retrieval grounds decisions in enterprise context. The asynchronous layer handles slower, complex decisions without leaving users waiting blindly. Human review and feedback keep the assistant aligned with real RAPID decisions and stakeholder corrections.

Models used: Claude 3.7 Sonnet · large language models

Scale

• RAPID-D uses four specialized agent roles in the decision workflow.
  “Agents: Specialized Decision Collaborators”article, stated
• The chosen model was Claude 3.7 Sonnet.
  “Claude 3.7 Sonnet was ultimately chosen for its strong balance of quality, stability, and reliability.”article, stated

Alternatives considered (roads not taken)

A monolithic single-answer AI assistant

Rejected because: Coinbase wanted modular agents that collaborate, debate, and synthesize information instead of one black-box response.

The initial single-agent prototype

Rejected because: Leadership feedback showed the need for a more dynamic, interactive, explainable multi-agent debate.

Synchronous-only decision processing

Rejected because: Some complex decisions need more time, so asynchronous architecture keeps users informed and returns a reasoned output later.

AI replacement of human oversight

Rejected because: RAPID-D is designed to augment human deciders with scalable diligence, not replace them.

Patterns

multi-agent-orchestration (orchestration)

RAPID-D divides decision support across Analyst, Seeker, Devil's Advocate, and Synthesizer agents that collaborate and debate before a recommendation reaches the human decider.

rag-retrieval (retrieval)

The Contextual Recommender generates critical questions and uses enterprise search to retrieve relevant internal knowledge.

tool-use-orchestration (orchestration)

The decision assistant uses a document-context retrieval tool that generates targeted questions and searches enterprise knowledge sources.

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

The final recommendation is for a human Decider, and comments or corrections from users and stakeholders are captured for adaptation.

eval-driven-development (evaluation)

Coinbase compared recommendations against real RAPID decider outcomes and used leadership feedback to evolve from a single agent to multi-agent debate.

Tradeoffs

Use specialized agents for separate decision roles.

Chose analyst, seeker, contrarian, and synthesizer agents over a single black-box assistant.

Rationale: Different agents expose facts, broader context, risks, and final synthesis, making the decision process more transparent and reproducible.

Add a contrarian role.

Chose an explicit Devil's Advocate agent over only recommendation-supporting analysis.

Rationale: The system needed to surface weaknesses, risks, unstated assumptions, and unintended consequences.

Use asynchronous architecture for complex decisions.

Chose async processing with user updates over forcing immediate synchronous answers.

Rationale: Some decisions require more time for retrieval and synthesis, but users still need to be kept informed.

Choose Claude 3.7 Sonnet.

Chose Claude 3.7 Sonnet over other leading model candidates.

Rationale: Coinbase selected it for the best balance of quality, stability, and reliability in benchmark review.

Gotchas (what bit them)

A single recommendation path can reinforce bias.

Root cause: Strategic decisions need adversarial challenge, not only support for the initial recommendation.

Resolution: Add a Contrarian Agent whose sole purpose is to argue against the initial recommendation and probe risks.

One-agent reasoning was not explainable or interactive enough for deciders.

Root cause: The initial prototype analyzed a document and presented reasoning, but feedback revealed the need for deeper explainability.

Resolution: Move to multi-agent debate with a synthesizer and asynchronous user experience.

Complex decisions may require longer-running analysis.

Root cause: Multi-agent retrieval, debate, and synthesis can take more time than a simple synchronous response.

Resolution: Implement asynchronous architecture so users stay informed and receive a well-reasoned output without unnecessary delays.

Evaluation

Coinbase evaluated RAPID-D through human review, comparing final recommendations against real RAPID decisions made by Coinbase Deciders. User and stakeholder comments are captured and analyzed against the assistant's original recommendation to improve later recommendations in the same flow.

Metrics: human review agreement with real RAPID Decider decisions · benchmark scores across leading models · quality/stability/reliability model selection · user or stakeholder corrections

Outcomes

• RAPID-D evolved from a single-agent prototype into a multi-agent debate and synthesis workflow
• the process became more transparent, consistent, and reproducible
• Claude 3.7 Sonnet was selected for quality, stability, and reliability
• the assistant augments rather than replaces human oversight

The source provides qualitative organizational impact rather than adoption counts; the strongest quantified details are the four-agent design and Claude 3.7 Sonnet model choice.