From Chatbot to Colleague: Coinbase's Vision for AI as a Teammate
What if an AI could do more than just answer questions? Coinbase's internal AI system, CEEcil, is built to act as a proactive digital team member with memory and judgment, moving beyond simple chatbots to become an integrated part of daily workflows.
Proactive Collaboration, Not Just Passive Responses
CEEcil's defining feature is its initiative. It can be @-mentioned in Slack, joining conversations like a human colleague. During production incidents, it actively monitors relevant channels, reacts with emojis, and follows up if issues stall. Its capability extends to safety judgments—in one instance, it refused to commit an operations document containing customer PII to Git, suggesting a sanitized version or compliant storage instead.
Architecture: Engineering Memory and Control
The system is powered by a Go service and an AI agent runtime, connected via the Model Context Protocol to tools for memory retrieval, knowledge bases, live queries, and Slack. Robust safeguards are in place:
- A real-time “kill switch” to halt the agent.
- Strict rate and spending limits on API calls.
- Mandatory human review and merge for all AI-generated code.
- Comprehensive audit logging for all actions.
The Three-Tier Memory System
CEEcil's “long-term memory” is central to its role as a teammate, enabled by a three-layer process:
- Continuous Extraction: It constantly pulls key data—decisions, blockers, owners—from authorized channels.
- Nightly Consolidation: Short-term observations are synthesized into structured long-term summaries.
- On-Demand Retrieval: When queried, it dynamically accesses historical memory, recent chats, and real-time data.
Layered Model Strategy: Balancing Cost and Intelligence
To optimize cost and latency, CEEcil uses a tiered approach. Simple queries are handled via APIs or knowledge bases. Only complex, multi-step tasks are routed to more powerful (and expensive) LLMs. This ensures efficiency for routine operations while reserving advanced reasoning for where it's needed most.
Currently, CEEcil assists with production incidents and provides expert knowledge that previously took hours to obtain. Coinbase envisions scaling this model into multiple team-specific AI agents, transforming AI from a tool that answers questions into a persistent participant that remembers, initiates, and gets real work done.