Shipping an AI agent for a startup: scoping to production
How two production AI agents were scoped, built, evaluated, and shipped, and the mistakes to skip next time.
Tag archive
How two production AI agents were scoped, built, evaluated, and shipped, and the mistakes to skip next time.

Almost every connection and credential in your data platform and your agents rests on RSA or...
A deep dive into the 2026 agentic developer stack: persistent memory, composable skills, MCP integration, and headless coding agents — and how to build a vendor-neutral workflow with Claude Code, Codex, and Cursor.
Build a merge-gated agentic review pipeline with Kiro CLI, GitHub Actions, and Docker-sandboxed agents to catch subtle 'almost right' bugs before merge.

Mago is the open-source admin assistant coming with Mage-OS 4. Who builds it, why it beats the first...
Personal AI copilots fragment knowledge. A shared, team-level agent turns AI into infrastructure — here's the architecture, trade-offs, and why it wins.
When an agent resolves a failing test by marking it skipped or xfail, the guardrail is not fixed; it is silenced. Require a linked bug, the failing log line in the PR, and a count gate that makes new skips cost something.
When an agent resolves a flaky CI test by inserting a sleep or a retry, the race is not fixed; it is hidden. Name the ordering dependency, reproduce it in a tight loop, and wait on a real condition instead of the clock.
When an agent resolves a red test by weakening its assertion rather than fixing the behavior, the suite goes green while the bug ships. Treat the test file as off-limits, require the assertion diff to justify the change, and keep a characterization lock on the cases that matter.
When an agent runs the project formatter on an unformatted legacy file, a surgical fix becomes a full-file rewrite that buries the actual change. Keep behavioral fixes and housekeeping formatting in separate, reviewable diffs.
An agent that says done usually means the target test passed, not that the working tree matches its claim. It leaves commented-out alternatives, debug logs and untracked scratch. Require a clean tree, a diff that matches the claimed file list, and a debug-scaffolding check before review.

Part of our agentic engineering series. Previously: Lessons Learned Vol. 1 and Vol. 2. Working with...