DEV, the community site that runs on the open-source Forem codebase, spent the last month pointing coding agents at its own maintenance backlog. The results are unusually specific for this genre: in a post published Monday, DEV co-founder Ben Halpern reports that average request time inside Rails fell from about 70 milliseconds to 16 milliseconds, measured across roughly 300,000 requests a day in Better Stack.

The supporting numbers move together, which is the part worth taking seriously. Database time per request went from 39ms to under 12ms. View rendering went from 12ms to 2.6ms. Redis time went from 7ms to 1.3ms. Before the work, P99 response time sat above one second much of the time; that tail is the thing users actually feel.

The mechanism was not one clever agent. Halpern describes a set of scheduled loops on a self-hosted OpenHands hub, driven through the Agent Client Protocol. Over 30 days, 475 pull requests merged into Forem. Since the hub went up on October 19, the split was 351 autonomous merges against 93 that he wrote or drove himself: 172 performance changes, 93 test-suite speedups, 32 hot paths rewritten in Rust through the magnus bindings, 29 dependency upgrades, 20 production error fixes and 5 issue-to-pull-request conversions. An earlier, fully interactive push in early October shipped 31 PRs and only got average latency from 70ms to 50ms.

Four configuration changes did most of the work, and each came from watching an earlier version fail.

**Telemetry instead of code reading.** The first `perf-loop` read Ruby files and guessed at bottlenecks, producing micro-optimizations that barely moved a graph. Wiring Better Stack's ClickHouse spans and a read-only Postgres gateway into each pre-flight snapshot — behind a privacy guard — forced every proposed change to justify itself against live sequential scans, unused indexes and per-endpoint compute.

**A router across model families.** The loops started pinned to one model. Quota ceilings and OAuth token-refresh collisions pushed the team to a three-provider router — Claude Opus 5.5, OpenAI's `codex-acp` (GPT-6.1 Sol) and Google's Gemini 4 Argon via the Antigravity CLI — rotating round-robin with an eight-hour cooldown cap, and switching model family whenever a PR repair fails.

**One worktree per loop.** Eight parallel loops in a shared checkout exposed every sharp edge in Git and Linux: `git gc` collisions on parallel fetches, a stale `.gitconfig.lock` after a crashed boot, zombie processes piling up until `tini` was added as PID 1. Each run now gets a detached-HEAD worktree with self-healing lock cleanup.

Specialised loops instead of one general one. `perf-loop` handles ActiveRecord queries, Redis caching and view partials; `test-loop` parses 30 Knapsack RSpec shards to trim FactoryBot churn; `rust-loop` is gated to one open PR at a time so precompiled binaries never collide, verified by 1,000-input differential specs; `error-loop` and `deps-loop` carry strict filters. Loops run on 10-minute daytime ticks (7am–7pm ET), quarter-cadence at weekends, with 5/10/20/40-minute exponential backoff on PR check-ins. Deterministic Python gates sit in front of every cron so an empty queue exits in two seconds and spends zero tokens. A reviewer on a different model family checks blast radius and auto-merges low-risk performance and test PRs once CI passes.

What this is not: a controlled benchmark. It is one operator, one codebase and self-reported numbers, with no external audit. The headline is also server-side request time inside Rails, not the full page load a browser experiences — network, assets and third-party scripts sit outside the 16ms.

The interesting claim is narrower than "agents write your code now". It is that the maintenance layer — query plans, test suites, dependency bumps, hot-path ports — is measurable enough that a loop can be graded on it automatically and rejected when the graph does not move. That is precisely the condition under which autonomy stops being a gamble and becomes a queue.