August 2026:
353 of 354 Releases, No Human

August was smaller than July, and we would rather say that plainly than bury it. 2,650 commits against July's 3,401 — down 22%. 672,000 lines added against 1.13 million — down 40%. 18 active projects against 26. Publishes to production fell from 490 to 354. Every one of those numbers went the wrong way, and we re-counted the org repository by repository through a second, independent code path before publishing them: 2,656 against 2,650, a 0.2% spread. The month really was smaller.

What went the other way is the number this report is named after. Of the 354 releases that reached production, 353 were shipped by the agent and exactly one by a person. In July that split was 476 of 490 — 97%, and we called it the number that mattered. In August the remaining 3% very nearly disappeared. That figure is measured, not modelled: every August merge commit whose subject begins Publish:, across all 70 repositories we touched, tallied by who authored it.

And the work got denser rather than smaller. August pushed 22.88 billion tokens through Claude — up 39% on July — while producing 22% fewer commits. That is roughly 8.6 million tokens behind every commit, against 4.8 million in July. Fewer, larger, more-reasoned units of work: more of the month spent deciding what to change, less of it spent typing.

If you want the honest explanation for why the commit count fell, it is this: most of August went into the parts of Fabric that nobody demos. A platform that ships 353 releases a month without supervision needs to be able to tell you when it is broken, and for most of the summer it could not do that reliably.

So August built the plumbing. The fleet now owns its own log sink — a drain receiver and a queryable store, so a failing agent produces a searchable record instead of silence. Alerts learned to clear themselves when the condition they fired on goes away, which sounds trivial and is the difference between an alert you trust and one you mute. A killed preview invocation now leaves a record behind. A transient database fault during provisioning requeues instead of failing the whole setup.

Jerry got narrower rather than broader, deliberately. He now reaches the client's own application and learns from the fixes we make; his web access is scoped to the client's own hosts; and one prompt-level capability claim that was simply false was removed. Refusals got more honest too — on the Velocify write path, does not exist and not authorised are now different answers instead of one vague failure.

The integration work continued underneath that: secure vault folders shareable with named client companies, outbound email for client apps without those apps ever holding the mail-provider key, argument validation against the vendor's own WSDL rather than echoing back whatever was sent, and scoping work on two new data exchanges. None of it makes a good screenshot. All of it is what the 353 rests on.

August pushed 22.88 billion tokens through Claude — up 39% on July even as commits fell — and 21.86 billion of them were read from cache (95.8%, our highest month yet). Almost all of that volume is 1M-token context read straight from cache rather than recomputed: written once, then re-read across sessions and across agents. That is what makes it affordable to spend more reasoning per change instead of more commits. Across the trailing twelve months we have now processed 72 billion tokens at ~95% cache efficiency.

The busiest repository in the organisation last month was a mortgage CRM and borrower portal. It took 710 commits in August. Not one of them was written by a human being: 332 came from the Fabric coding agent, 330 from Jerry, and 48 from Neo, our internal engineering agent.

The commit log is more revealing than the total. 234 of those 710 commits carry the subject “apply requested change”, and 178 are publishes to production. That is the entire loop, visible in the history: a client describes something in plain English, Jerry applies it, and it goes live. Two hundred and thirty-four times in a month, on one product, with no ticket, no spec, and no engineer in between.

This is what we meant in July when we said the person steering the software is now the person who wanted it. August is the first month where you can read that claim straight off the commit history of a single repository rather than infer it from an aggregate.

Iterative refinement is the largest bar again, and it still is not a bug count — with AI the first pass is a starting point, so most “fixes” are refinements. What changed this month is the ratio at the end of that loop. 353 of 354 production releases went out with nobody watching, against 476 of 490 in July. The loop did not get smaller; the last person standing in it stepped out.

• →353 of 354 production releases shipped by the agent — a single release all month had a human on it
• →Fleet log sink — the platform now receives, stores and queries its own logs, so a failing agent leaves evidence instead of silence
• →Self-clearing alerts — a condition that resolves now retires its own alert, closing the trap where every alert eventually gets muted
• →Remote workspaces — proof-of-concept for browser-delivered dev environments on short-lived, single-use, identity-bound tokens
• →Secure vault — folders shareable with named client companies rather than all-or-nothing
• →Outbound email for client apps without the client app ever holding the mail-provider credential
• →Vendor call validation against the provider's own WSDL, and refusals that distinguish 'does not exist' from 'not authorised'
• →Provisioning resilience — a transient database fault requeues instead of abandoning a half-built project
• →Jerry scoped tighter: reaches the client's own app, learns from our fixes, and one false capability claim removed from his prompt
• →A greenfield internal tool went from empty repository to production release in under 12 hours

• ✓Append-only guarantees on consent records narrowed so they protect the data without blocking legitimate identity re-keys
• ✓Foreign-key cascade rules re-swept across a production schema after 27 new keys had drifted in with the wrong behaviour
• ✓Migration-drift detection taught to ignore comment-only edits, and to ring a staff alert when code runs ahead of any tier's schema
• ✓Row-level security applied to internal migration bookkeeping tables
• ✓Cross-project agent handoffs now mount the source project's preview rather than its last published build
• ✓Publish semantics clarified per-turn, so a turn that ends in a question no longer ships

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

Activity spread across 18 active repositories out of 230 total in our organization. Ranked by commit volume. Projects are anonymized — the description says what kind of product it is, never whose:

A down month is a useful thing to publish, because it tells you which of your numbers are load-bearing. Volume fell 22% and the thing we actually sell — software reaching production without us — got better. What we are watching next:

• The large greenfield build we trailed in the July report did not land in August; it is still coming, and we would rather say so than pad this report with it
• Holding the publish ratio where it is while volume climbs back — 353 of 354 is easier at 2,650 commits than at 4,000
• Turning August's log sink and self-clearing alerts into the diagnostics customers see, not just the ones we see
• Onboarding more customers onto Fabric and continuing to take it to market
• Deepening mortgage integrations — origination, pricing, credit, title, voice, and messaging