·3 min read
Shipping an AI agent in two weeks
A no-drama playbook for getting a useful AI agent into production fast — scope, guardrails, and the boring infra around it.
AI agentsshipping
Pick one job, not a personality
An agent that "helps with everything" never ships. Pick one repeatable task — qualifying leads, drafting replies, summarising calls — and design around that.
What we wire up
- A single tool the agent is allowed to call
- A typed input + output schema
- Logs of every prompt, tool call, and result
- A kill switch the operator actually trusts
Why two weeks is realistic
Week one is scope, evals, and the happy path. Week two is the boring edges: timeouts, retries, fallbacks, and a thin UI for humans to take over.
Ship that, watch the logs for a week, then add more tools.
