Writing on agents in production.
What we've learned from shipping. No takes on whether agents are real.
// featured post
Why your eval suite belongs in CI.
If your evals don't run on every PR, they're not evals — they're a screenshot from a meeting in March.
Read · Engineering team
A2A and MCP: pick both.
They solve adjacent problems. The mistake is treating them as a fork in the road.
The runbook that ships with every agent.
Failure modes, escalations, rollback conditions. Boring on purpose.
Durable workflows are not retries.
A reasonable mental model for when you actually need a durable orchestrator vs. just better backoff.
Why we run on-call for 60 days after handoff.
Because the first 60 days surface every assumption the eval suite missed.
How to build a retrieval eval set in a week.
A practical recipe. 200 questions, your real corpus, baselines, and a "do not optimize this" set.
START
Ship the first system.
Fixed-price discovery in 2 weeks. You leave with an architecture, a working spike, and a build plan.