Finance / Crypto exchange
Coinbase: Reviewing internal decision documents
- Company
- Coinbase
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
According to a Coinbase engineering blog post, Coinbase uses the RAPID (Recommender, Agree, Perform, Input, Decider) framework for critical decisions in which input is gathered from key stakeholders across functions and an accountable Decider makes the final decision. A RAPID document contains the recommendation with its details, pros and cons, risks and benefits, and whether the decision can be reversed, and each named person can add an Agree or Disagree with additional context. The post said the company added AI to this framework to systematically surface unseen risks, reduce cognitive bias and provide a transparent, auditable layer of analysis for its most critical decisions.
Technology and data
According to the same post, Coinbase built RAPID-D, an internal decision support tool that augments the RAPID framework. Instead of one AI producing a single answer, RAPID-D has four agents with different roles working together. The first agent reviews only the RAPID document and gives a baseline recommendation based on the facts and arguments in it. The second agent generates key questions about the document in areas such as security, market impact, cost, user experience and scalability, then uses the enterprise search tool to find answers in Coinbase's internal knowledge and gives a recommendation that reflects the wider organizational context. The third agent builds the strongest case against the initial recommendation to find weaknesses, unstated assumptions and risks. The last agent compares the three results and gives the human Decider a final recommendation with an explanation of its reasoning and trade-offs. According to the architecture diagram in the post, the RAPID document is loaded through a Document fetch API and the final output is a CB-GPT recommendation. Comments made by the user during a session, or later by stakeholders in the RAPID document, are compared with the original recommendation and used in the next recommendations within the same decision flow.
Results
According to the same post, Coinbase measured accuracy through human review, comparing each final recommendation from RAPID-D with the actual decision documented by Coinbase's RAPID Deciders. According to the accuracy table image in the post, Claude 3.7 Sonnet scored 86%, Gemini 2.5 pro 85% and GPT-4o 71%, and the company chose Claude 3.7 Sonnet for its balance of quality, stability and reliability. The post explained that the work started with a single agent that analyzed a document and presented its reasoning, with early feedback from leaders gathered manually, and that the current version adds debate between agents and deeper explanations and handles long-running decisions asynchronously while keeping users informed. The post said this step-by-step structure makes the whole decision process transparent, consistent and reproducible.
Limits and open questions
The post did not disclose how many decision documents were used in the accuracy evaluation, the evaluation period, the number of employees using RAPID-D or the number of decisions it handled. This research could not confirm the operating status after the post was published from official sources.
Sources
- Making Smarter Decisions, Faster with AI at Coinbasecoinbase.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.