One sentence in, one merged PR out.
Write-ups, with receipts, from the team building RunAI Coder — an autonomous coding agent.
RunAI Coder is an autonomous coding agent: you hand it a goal in natural language, it plans, fans the work out to parallel workers, runs the tests, and files the result as a delivery for your sign-off — with the diff, the test run and the full step-by-step trace attached. The unit of work isn't a suggestion; it's a merged PR.
Everything in this repo follows the same rule as the product: claims ship with receipts. Posts link to measured data, real runs and reproducible methodology, and we state limitations ourselves rather than waiting for someone else to.
A public SWE task on pallets/flask (CLI feature + tests, all green), recorded from a production session — playback sped up 8×, content untouched. The money-bag indicator at the bottom left is the live compression savings multiple climbing as cached and compressed context accumulates:
| Metric | Value | Scope |
|---|---|---|
| Merged squash PRs in 7 days | 968 | 2026-07-14 → 07-21 PDT, one developer, one machine |
| First-pass completion | 98.8% (1,786 / 1,807 deliveries) | same window |
| Best single day | 98 merged PRs | 2026-07-21 |
| Context compression | 2.83× median (p10–p90 2.26–3.08×) | n = 7,222 production requests |
| Effective input price | ≈ $0.26 / M tokens (≈ 2.6% of list) | measured account, caveats published |
Every number traces back to a machine-recorded ledger or a reproducible script — formulas, sample windows and known limitations are in the performance report.
| # | Title | Date |
|---|---|---|
| 001 | One sentence in, one merged PR out — what RunAI Coder is, a week of dogfooding numbers with the caveats attached, and why a merged PR costs $10 | 2026-07-30 |
| 002 | Our house rules for letting an AI agent commit code — five boring-on-purpose guardrails for giving an agent write access, and what they don't solve | 2026-07-31 |
| 005 | The bill is 99% input tokens: cost engineering for a coding agent — one ledger day, three levers from $10/M to an effective $0.26/M, and the SWE-bench A/B that keeps it honest | 2026-08-03 |
| 006 | Why your coding agent forgets: a field guide to context rot — the two-budget model of a context window, what compaction quietly costs, and what helps if you just use these tools | 2026-08-04 |
| 007 | 33k tokens before hello: the anatomy of your coding agent's preamble — what the viral token-overhead study really found: configuration swamps defaults, cache churn beats size, and a two-minute way to measure your own | 2026-08-05 |
| 008 | Nobody reads your architecture overview. Including the agent. — why an instruction file is rent, not documentation: three conflicting evals reconciled, a five-lane probe on Flask, and the harness that never read our AGENTS.md at all | 2026-08-06 |
Terminal apps for Windows (10/11 x64), macOS (Apple Silicon) and Linux (x64), CLI included — an Android remote-control app is rolling out, iOS next. Free starter pack: US$25–30 of value, one-time, 14 days, no card required.
Text and images in this repository: CC BY 4.0 unless noted otherwise. Product screenshots, recordings and brand assets belong to RunAI Coder.