Coinbase CEEcil Support Assistant
Routes simple queries first, delegating only open-ended issues to the full agent. Task: Coinbase's CEEcil support assistant uses a routing architecture where a small, cheap model classifies incoming customer queries. Simple lookups are handled by deterministic code paths, while only open-ended or complex issues are delegated to a full LLM agent. The agent can generate Pull Requests for fixes but requires human approval to merge them.; tools: Tool described in the source. Human role: The source summary does not fully specify human review and operational roles; actual deployment should include human oversight based on task requirements.; limitations: Information is derived from a secondary summary of a public engineering blog, which may contain marketing bias. The text does not specify the exact model sizes, infrastructure costs, or the volume of tickets handled. Human oversight is mandatory for merging code changes and handling high-risk escalations..
Tools
Tool described in the source
Published on Agent Work Atlas:
Overview
Routes simple queries first, delegating only open-ended issues to the full agent.
Workflow
Coinbase's CEEcil support assistant uses a routing architecture where a small, cheap model classifies incoming customer queries. Simple lookups are handled by deterministic code paths, while only open-ended or complex issues are delegated to a full LLM agent. The agent can generate Pull Requests for fixes but requires human approval to merge them.
Reported result
The source describes the architectural design and capabilities of the CEEcil assistant, noting that deterministic handlers respond in seconds. However, specific performance metrics, resolution rates, or cost savings are not provided in the text. Independent validation of these operational outcomes is required.
Human role
The source summary does not fully specify human review and operational roles; actual deployment should include human oversight based on task requirements.
Limitations
Information is derived from a secondary summary of a public engineering blog, which may contain marketing bias. The text does not specify the exact model sizes, infrastructure costs, or the volume of tickets handled. Human oversight is mandatory for merging code changes and handling high-risk escalations.
Try it
Act as a technical lead implementing a customer support routing system similar to Coinbase's CEEcil. 1. **Context**: We have a backlog of support tickets [replaceable input: e.g., 'billing', 'technical']. 2. **Task**: Design a two-tier routing workflow. - Tier 1: Use a lightweight classifier to identify simple queries (e.g., password reset, status check) that can be answered by deterministic scripts or FAQ retrieval. - Tier 2: Route complex, open-ended issues to a full LLM agent with access to [replaceable input: e.g., 'CRM API', 'Knowledge Base']. 3. **Constraints**: The agent must never auto-merge code fixes; it should only draft Pull Requests for human review. 4. **Deliverables**: - A Python script skeleton for the router using [replaceable input: e.g., 'LangChain', 'LlamaIndex']. - A prompt template for the Tier 2 agent. - A dry-run plan: Simulate 10 sample tickets [replaceable input: 'ticket_logs.csv'] to verify routing accuracy before production deployment. 5. **Success Metric**: 80% of simple tickets resolved without LLM inference; 0% unauthorized merges. Ask clarifying questions if the ticket categories or available APIs are unclear.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.