What I've learned shipping agents.
Mostly about the unglamorous half of agentic AI: what breaks, what the demos skip, and what it costs. Written from things I actually ran.
Your product is rarely one repository. Why should your agent's memory be? →
A coding session in one repo works out that the bug is actually in the SDK. The session ends and the finding is gone. What building a local, provenance-tracked memory layer across repos forced me to get right, including the bug that broke the exact case the tool exists for.
The context-switching tax on running multiple agents →
Running Claude Code, Gemini CLI and Codex across separate terminals cost more than the ten seconds it looked like. What changed when every session lived in one place with its own isolated worktree.
Who verifies the verifier? →
Five models built a Notion clone unattended with permissions enforcing separation of duties. The defect ledger looked great, then a passing E2E test turned out to be logging that the drag didn't work. What that hole says about multi-agent systems.
Coding interviews are testing the wrong thing →
AI writes most production code now, but interviews still ask whether you can hand-write a binary tree traversal in 45 minutes. Four questions that would actually tell you something about a candidate.
Treat your AI agent like infrastructure, or don't run it →
Everyone's installing autonomous agents with shell access on their personal laptop, where their SSH keys and browser sessions live. The minimum viable hardening checklist before you give an agent real system access.