Coinbase has disclosed CEEcil, an internal AI agent designed to function like a team member rather than a basic chatbot. It can be mentioned in Slack, maintain long-term memory, take initiative and make certain operational judgments. A three-layer memory system extracts authorized-channel context such as decisions, blockers and owners, converts nightly observations into durable summaries, and retrieves relevant history, recent discussions and live data when needed. CEEcil uses a tiered model architecture: simple requests are handled through APIs or a knowledge base, while complex, multi-step work is routed to more capable AI models to reduce cost and latency. The system can join relevant Slack conversations, react with emoji and follow up when no one handles a question. Coinbase said CEEcil once refused to commit an operations document containing customer identity information to Git, recommending a redacted version or compliant storage instead. Built from Go services and an AI agent runtime, CEEcil connects memory retrieval, knowledge-base access, real-time queries and Slack through MCP (a protocol for connecting AI systems with tools). Coinbase has also added a real-time kill switch, limits on calls and spending, auditing, and mandatory human review and merging of generated code. The company said CEEcil already helps teams handle production incidents and obtain expert knowledge that previously could take hours to access. Coinbase plans to extend the model to multiple AI agents serving specific teams.