Engineering, orchestration, and what we learn running the pipeline on our own work. Short. Specific. The kind of writing we’d want to read.
A PR that opened, conflicted and closed forever. 84 failed tasks in three minutes on one worktree error. A build that failed the same way every cycle. Three layers of Colony hit the same loop, and each one got a breaker with a manual way out.
Read the postColony decides what to retry by matching error strings, and every string-matcher eventually matches the wrong string. Four misclassifications, where they were fixed, and the negative tests that keep the predicates honest.
Colony requires a `{file, line, content}` citation for every claim that a review item is already correct, then replays it against the actual workspace before a PR can pass. The fabrication incident behind the design, the replay mechanism, and the hard override on a convincing but false verdict.
A companion to "What goes wrong with autonomous coding agents in production" — six incident classes from the granular log, each grounded in PR numbers, covering ENFILE cascade, VM freeze, Marshal force-push recovery, epic chain depth cap, flag/state/task desync, and fabricated evidence at the trust boundary.
1,232 PRs, $3.70 median cost, 18-minute median active dev time. The numbers for July 2026, agent by agent.
913 PRs, $3.50 median cost, 23-minute median active dev time. The numbers for June 2026, agent by agent.
Across 1,284 PRs over 30 days, Builder took 53.6% of the total spend. Here's why that's the right allocation, what two context-budget experiments changed in the pipeline, and the turn-limit approach that didn't work as the primary termination signal.
Colony's pipeline used to treat GitHub labels as the authority for issue state. Running multiple workers against an eventually-consistent API produced a race that label writes alone could not fix — here's the incident, the migration to Postgres-first transitions, and what atomic state changes enable.
May numbers by agent, corrected: 899 PRs, $3.07 median cost, 27-minute median active dev time.
This isn't an announcement. It's a disclosure of evidence already accumulated.
Six failure patterns we encoded as recovery logic — and why they are hard to see in a demo.
A PR that opened, conflicted and closed forever. 84 failed tasks in three minutes on one worktree error. A build that failed the same way every cycle. Three layers of Colony hit the same loop, and each one got a breaker with a manual way out.
Read the postColony decides what to retry by matching error strings, and every string-matcher eventually matches the wrong string. Four misclassifications, where they were fixed, and the negative tests that keep the predicates honest.
Colony requires a `{file, line, content}` citation for every claim that a review item is already correct, then replays it against the actual workspace before a PR can pass. The fabrication incident behind the design, the replay mechanism, and the hard override on a convincing but false verdict.
A companion to "What goes wrong with autonomous coding agents in production" — six incident classes from the granular log, each grounded in PR numbers, covering ENFILE cascade, VM freeze, Marshal force-push recovery, epic chain depth cap, flag/state/task desync, and fabricated evidence at the trust boundary.
Across 1,284 PRs over 30 days, Builder took 53.6% of the total spend. Here's why that's the right allocation, what two context-budget experiments changed in the pipeline, and the turn-limit approach that didn't work as the primary termination signal.
Colony's pipeline used to treat GitHub labels as the authority for issue state. Running multiple workers against an eventually-consistent API produced a race that label writes alone could not fix — here's the incident, the migration to Postgres-first transitions, and what atomic state changes enable.
Six failure patterns we encoded as recovery logic — and why they are hard to see in a demo.
1,232 PRs, $3.70 median cost, 18-minute median active dev time. The numbers for July 2026, agent by agent.
Read the post913 PRs, $3.50 median cost, 23-minute median active dev time. The numbers for June 2026, agent by agent.
May numbers by agent, corrected: 899 PRs, $3.07 median cost, 27-minute median active dev time.
This isn't an announcement. It's a disclosure of evidence already accumulated.
No posts in this category yet.
