Coinbase has disclosed the architecture and internal use cases of CEEcil, an in-house AI system designed to function more like a team member than a basic chatbot. The company said CEEcil can be mentioned in Slack, keeps long-term memory, follows up proactively, and can make certain judgments.
Three-layer memory system
According to Coinbase, CEEcil preserves team context through a three-layer memory structure. In the background, it continuously extracts decisions, blockers, owners and other information from authorized channels. Each night, it consolidates short-term observations into long-term summaries. When answering a question, it then retrieves historical memory, recent conversations and real-time data as needed.
Tiered model architecture to manage cost and latency
CEEcil also uses a layered model architecture. Simple queries are handled first through APIs or knowledge bases, while only complex, multi-step tasks are routed to stronger AI models. Coinbase said this approach is intended to reduce both cost and latency.
Can join discussions and follow up on unattended issues
In practical use, CEEcil can join relevant Slack discussions on its own, respond with emoji, and automatically follow up when a question goes unanswered. Coinbase said the system is already helping teams handle production incidents and answer expert-level questions that previously could take hours to resolve.
Refused to submit a document containing customer identity information
Coinbase said CEEcil once refused to submit an operations document containing customer identity information to Git. Instead, it suggested using a redacted version or a compliant storage method, which the company cited as an example of its security judgment.
Built on Go services and an AI agent runtime
From an infrastructure standpoint, CEEcil is composed of Go services and an AI agent runtime, and connects through MCP to memory retrieval, knowledge bases, real-time query systems and Slack. Coinbase said it has also put in place a real-time kill switch, limits on call volume and spending, audit mechanisms, and a requirement that any code generated by the system be reviewed and merged by humans.
Coinbase wants to expand the model to more specialized agents
Looking ahead, Coinbase said it hopes to extend this model into multiple AI agents built for specific teams, so AI can do more than answer questions and instead retain memory, participate proactively and complete real work.