Coding agents,
in practice
Field notes from the team building Farfield. What works in real codebases, what breaks in prod, and what we are still trying to figure out.
Coding agents in Slack make the whole company an apprentice
Shopify, Stripe, Ramp, and WorkOS each built the same thing in eighteen months: a coding agent that lives in a public Slack channel. The team trains the agent. The team also trains itself, by watching the agent work. This is what AI engineering looks like when it stops being personal.
Hermes Agent: one Python loop, thirty-plus platforms, no plugin manifest
Reading Hermes Agent: one ReAct loop, thirty-plus chat platforms, eight sandbox backends, an SQLite session store. Notes from two days in the codebase.
Async coding agents need async reviewers. Slack is the obvious answer.
AI writes more code than humans can review synchronously, background agents are a real category, and Slack is already the coordination layer. The winning shape is a Slack-native background quality agent.
If your coding agent doesn't reduce your maintenance bill, it's costing you twice
AI coding agents that double output without subtracting maintenance cost are net-negative. The honest productivity equation, the asymmetry that breaks it, and what to measure instead.