September 2026:
4,146 Commits — Our Biggest Month

August was the down month we published on purpose; September is the answer. 4,146 commits — up 56% on August and 22% above July's previous record of 3,401 — across 26 active projects, back to July's count after August's 18. 504 releases reached production, up from 354, the most we have ever shipped in a month. We re-counted the organisation repository by repository through a second, independent code path before writing any of this: 4,146 against 4,146. No spread at all.

The autonomy number moved the other way, and we would rather say so than round it. Of the 504 releases, 500 were shipped by an agent and four by a person — 99.2%. August was 353 of 354. A single human release became four; the ratio slipped by half a point while volume rose by more than half. It is measured the same way as before: every September merge commit whose subject begins Publish:, across all the repositories we touched, tallied by who authored it.

And the work stayed dense. September pushed 33.65 billion tokens through Claude — up 47% on August, roughly double July — at 95.1% cache, which is about 8.1 million tokens behind every commit. August was 8.6 million, July 4.8. Commits grew by 56% and tokens by 47%, so the reasoning per unit of change barely moved while the number of units took off. We added 1.79 million lines by the same method as every other month; roughly a fifth of that is generated API specifications and starter-template scaffolding, so read it as about 1.4 million lines of written code, not 1.8.

The single busiest repository in the organisation last month was Fabric itself: 1,330 commits, with the platform moving from version 0.344 to 0.607 in thirty days. If August was the month we built the plumbing that tells us when an agent is broken, September was the month the fleet started to act on it.

Our internal engineering agent, Neo, got a rebuilt harness. Its prompts are now signed documents rather than text baked into a runner, so what an agent was told is something you can read, version and diff. Agents now come in kinds — a coordinator that fans a large story out to children and gathers the result, builders that do the work, a reviewer and a verifier that judge it — and each kind is tiered to the model that suits the job: a fast, inexpensive model for the doers, the strongest one for the judge. A daily live probe now runs every pinned agent SDK against every model we use, and rings a staff alert when one stops working, instead of letting a model upgrade quietly break a turn halfway through.

Landing code got safer in the places a fleet is most likely to hurt itself. A rebase conflict is now named — the files, and the commits on the other side — before a landing is attempted rather than after it fails. A conflict in release notes alone no longer loses a landing; the resolver that was supposed to handle it had never worked against the real file, and now it does. A pipeline write now passes a user-acceptance gate at the single choke point every write goes through, and a staff kill switch for the architect agent sits in the portal beside the host-level one.

Jerry kept getting more reliable at the thing clients actually see. His prompt became a signed, rendered document; his event tokens now outlive the longest turn; a class of mid-turn failures on the newest models was fixed by a pin update; and resumed conversations no longer lose their sandbox. The same month, an integration standard for messaging grew into a twelve-capability specification, and the two data exchanges scoped in August had their vendors' API specifications written down from the vendors' own documentation before a line of integration code.

September pushed 33.65 billion tokens through Claude — up 47% on August, and our heaviest month ever — and 31.91 billion of them were read from cache (95.1%, a shade under August's 95.8%). Output reached 100.7 million tokens, the first month above 100 million. Almost all of the input is 1M-token context read straight from cache rather than recomputed: written once, then re-read across sessions and across agents. Across the trailing twelve months we have now processed 105 billion tokens at ~95% cache efficiency.

Across the four busiest client-facing repositories, 525 commits carry the subject “apply requested change”. Every one was authored by the agent. Each is the same loop made visible in a commit history: someone describes a change in plain English, the agent applies it, and it goes live. 300 of them landed in one mortgage CRM and borrower portal alone.

The clearest case is a repository that did not exist on 1 September. It was created on the 15th. By the 30th it held 125 authored commits, 121 of them “apply requested change” — a new client product built almost entirely from requests, with four publishes to production along the way, in fifteen days.

This is the claim from July and August, now readable straight off a repository that is only two weeks old: the person steering the software is the person who wanted it, and the history shows no one in between.

Iterative refinement is a large bar again, and it still is not a bug count — with AI the first pass is a starting point, so most “fixes” are refinements. 500 of 504 production releases went out with no human on them. August was 353 of 354; the honest reading is that autonomy held at roughly 99% while volume rose by more than half, and that four people still chose to press the button themselves.

• →504 releases to production — the most in any month — 500 of them shipped by an agent and four by a person
• →Agent kinds and coordinators — a coordinator persona fans a story out to children and gathers the result, with builders, a reviewer and a verifier each tiered to the right model
• →Signed prompt documents — what an agent was told is now a versioned file you can read and diff, including Jerry's own prompt
• →Daily live SDK-by-model probe — every pinned agent SDK is exercised against every model, with a staff alert when one stops working
• →Safer landings — rebase conflicts named before they bite, and a release-notes-only conflict no longer loses a landing
• →User-acceptance gate at the pipeline's single write choke point, plus a portal kill switch for the architect agent
• →A greenfield client product from empty repository to 125 commits in fifteen days, 121 of them requested changes
• →Vendor API specifications written from the vendors' own documentation before any integration code — one an OpenAPI document of 170 operations — and a contact-data API that audits every read (the fact of it, never the contents)
• →A twelve-capability messaging specification, including a refusal that changes channel rather than failing quietly
• →The agent fleet on a phone — the Matrix console rebuilt for small screens, with search across every past session

• ✓An intermittent build-gate failure root-caused to a kernel memory-map limit rather than heap size
• ✓A daily timer on the runner host that checks a fresh export of the preview branch and alerts when the check did not run, not only when it failed
• ✓Mobile-first Matrix console, and session search that reaches past the most recent twenty
• ✓Agent transcripts that name the person who typed each prompt, so a transcript is attributable
• ✓A client order portal gained seller-field ordering, deed reporting, three lien types and lender mapping templates
• ✓A mortgage dashboard no longer reports earnest money as a cash shortfall

Files modified this month, grouped by area of the stack:

Activity spread across 26 active repositories out of 296 total in our organization. Ranked by commit volume. Projects are anonymized — the description says what kind of product it is, never whose:

A record month makes the next one harder to read honestly, so here is what we are watching rather than promising:

• Whether 4,146 is the new floor or a peak — September had a single day of 379 commits, and the daily average was 138
• The human-shipped releases: four is a small number, but it moved the wrong way, and we want to know why those four were pressed by hand
• Whether a rising token volume keeps its 95% cache rate — it dipped slightly this month as the base nearly doubled
• More customers shipping their own software on Fabric, and more products going from an empty repository to live in days, not quarters
• Deepening mortgage integrations — origination, pricing, credit, title, voice, and messaging