Coinbase builds AI's "Digital Teammate" CEOcil: Memory Team Context and Active Mission

On August 29th, Coinbase published the structure and application of its internal AI system, CEOl, which is not a simple chat robot, but is designed as an AI agent similar to a “team member” and can be @ mentioned in Slack with long-term memory, proactive follow-up and judgement. CEOil is now able to help teams deal with production accidents and to respond to expert knowledge that would have to wait hours to obtain; it is hoped that this model will be extended to several AI agents for specific teams, so that AI will not only be able to answer questions, but will also be able to continue to remember, actively participate and do its work. In practical applications, CEO will be able to take the initiative to join the Slack discussion, send an expression response and follow up when the issue is not addressed. Coinbase stated that the system had also refused to submit to Git operational files containing customer identity information and suggested that it should be replaced by dissensitive versions or compliance storage, reflecting its security judgement capability. In architecture, CEOil consists of Go services and AI agents running while they are connected to MCP tools such as memory retrieval, knowledge base, real-time queries and Slack. Coinbase has also put in place real-time “kill switches”, calls and spending caps, audit mechanisms, and requires humans to review and consolidate the code they generate. CEPil maintains a team profile through a three-tier memory mechanism: 1; continuous back-office extraction of information on decision-making, blocking matters and responsible persons in authorized channels; 2; consolidation of short-term observations into long-term summaries every night; 3; response to questions and retrieving, as needed, historical memory, recent dialogue and real-time data, using a stratification model structure, with simple queries being processed first and foremost through API or the knowledge base, with only more complex, multi-step tasks being assigned to a stronger AI model to reduce costs and delays。