Final checklist

0 of 21 done

• Direct several coding agents in parallel and ship reviewed, tested work every day.
• Configure agent tools for a team: context files, skills, hooks, MCP servers, and permissions.
• Read Python and TypeScript fluently, and C# where needed, and catch subtle bugs in agent diffs.
• Design the checking for agent code: practical tests, reviewer agents on other models, and deep reading only where mistakes cost the most.
• Explain async code and concurrency, and debug blocked workloads from evidence.
• Estimate and monitor the LLM cost of each task, and the coding-agent cost of each engineer.
• Route tasks across models based on measured trade-offs.
• Build and evaluate hybrid retrieval that respects access control, and know when tool-based search is better.
• Design MCP tool APIs with strict validation and authorisation, and distribute them to teams.
• Model agent workflows as durable state machines with human approvals, and run them on a durable execution engine.
• Deploy AWS infrastructure as code, and use queues, retries, dead-letter queues, and idempotency correctly.
• Instrument model, retrieval, and tool spans with OpenTelemetry, and go from an alert to the failing trace.
• Keep versioned eval datasets and CI regression gates.
• Define SLOs and error budgets for completed tasks, and stop runaway spend with hard budgets.
• Design what each workflow does when a model provider fails.
• Build threat models for prompt injection, tool poisoning, data exfiltration, and coding-agent permissions.
• Sandbox untrusted code.
• Provide platform SDKs, templates, and a shared agent setup with safe defaults.
• Lead an adoption effort and measure it honestly.
• Write clear RFCs, decision records, SLOs, runbooks, and postmortems, and lead an incident.
• Show a production product with real users and a working capstone platform that covers all of the above.