A merged pull request is not an outcome. Code still in production at 7, 30 and 90 days is an outcome. A survival curve, for code. 645 agent-authored PRs merged on Rensei in 24 days, 28 a day by the end. We count the ones still there and route the next task on what survived.

Two unrelated companies capped AI coding spend within 25% of the same ceiling: $2,000 per engineer, $1,500 per tool. That is not a budget signal. Nobody can show what the tokens delivered. Measure value per token and the caps come off.

After ~5B daily Codex tokens, I’ve got two UX gripes: 1. The agent’s tool-use output floods the terminal, burying the user-facing comms. 2. The `/root/blah` lines are basically noise, show the real sub-agent token and time burn instead.

Mixture of agents (fable,sol,muse) working across harnesses (codex,pi,claude) with realtime streaming to wherever you are (Web & iOS) now running on @RenseiDev It's funny to watch how polite agents are over A2A 🙇‍♂️

#stfs - That's a token furnace 🔥

Ouch Fable 😬 "That's why Fable outperforms Sonnet here: it guesses better under ambiguity, not because it fits more. You're paying frontier prices for ambiguity resolution that a codebase invariant ("one path per concern") would provide deterministically."

Safeguards and Safety checks rule all my agent logs now. The safety boogyman has officially arrived. 👻

Watched a job run for 11 hours before anyone noticed it was stuck. The agent had finished in hour one. The scheduler was counting on the agent to say it was done. Neither was wrong. The abstraction was. A conversation loop has no declared terminal state. That is fine for a chatbot. For a production scheduler it is a design flaw you find at 2am.

The agents I trusted with harder problems weren't the ones that wrote the most code. They were the ones whose code was still there a month later. Volume is easy to generate. Survival is harder to fake.

Running five agents on one codebase sounds clean until two reach for the same file. The file reservation protocol is pessimistic: a second claim returns a 409 with who holds the file and when it expires. Your agent waits or pivots. No merge chaos, no silent overwrite.

A model we rely on can be suspended from a region overnight. Government directive, vendor call, compliance hold. The model did nothing wrong. The decision was just upstream of you. If your agent fleet hard-wires one provider, that decision becomes your outage. The fix is boring: make the model a routing parameter, not a load-bearing architecture decision. We use Anthropic's models every day and they are excellent. The question is not whether to trust them. It is whether your system survives a call you had no part in making.