July 2026:
3,401 Commits — and 490 Trips to Production

July 2026 set another high. We shipped 3,401 commits across 26 active projects — up 36% over June — adding an estimated 1.13 million lines across 16,612 files, at an average of 110 commits per day. A team of three engineers, working alongside a fleet of AI agents.

The number we care about most isn't commits, though. It's 490 publishes to production — each one a preview build promoted to a live site. In June that number was 51. Before June it was zero, because the one-button publish flow didn't exist yet.

And here is the part that changed how we work: of those 490 production releases, 476 were shipped by Jerry — not by an engineer. Fourteen were a human merging by hand. That is 97% of everything that reached production going out without one of us in the loop. For two years the bottleneck on AI-built software was a person to review and release it. In July that stopped being true.

June introduced Fabric. July was the month it got used in anger — and most of our engineering went into the unglamorous work that makes a platform trustworthy: preview environments that sign you in reliably, spend controls that explain themselves instead of going quiet, provisioning that can't leave a project half-built, and diagnostics that say something when a page loads clean.

That work is what the 97% rests on. Jerry only ships unattended if the rails underneath him hold — so when a customer says “every issue I flag is already fixed by the time I check back,” what they're describing is a build that ran, a test suite that passed, a preview that came up, and a release that went out, with nobody from Mace watching any of it.

The other half of the month went into integrations — the reason a mortgage business can build something real on Fabric instead of a toy. Clients keep their own vendor relationships and hand over an API key once; the key lives server-side and is never held by the client's own application, so it can't leak from their code or a build log. July added Experian (soft-prequal credit, with the raw bureau report decoded into plain language) and ResWare (title and closing-fee quotes, dry-run by default so a quote can't bill by accident) to a platform that already covered origination, pricing, HELOC, voice and messaging.

The pattern held across every one of those: find the real cause, at the byte level, and fix that — not the symptom above it. Several of July's fixes were things that had been mis-diagnosed for days by looking at the wrong layer. We write those up internally so the next one takes minutes instead of an afternoon.

I've put about 25 hours into it and it's working almost perfectly. Every issue I flag, it's already fixed by the time I check back. I'm going to be using this a lot — for a long time.

“If we're going to take over the mortgage world, it's going to take a lot.”

— John, early Fabric customer (mortgage)

July pushed 16.43 billion tokens through Claude — a 50% jump over June — and 15.39 billion of them were read from cache (94.0%). Most of that volume is 1M-token context read straight from cache rather than recomputed: large contexts written once, then re-read across sessions. That efficiency is what lets output scale without a proportional jump in cost. Across the trailing twelve months we've now processed 51 billion tokens at ~94% cache efficiency.

For two years this section described our process: Mace engineers co-developing with AI instead of working from detailed specifications, iterating fast because the first implementation is a starting point rather than a final product. That was true, and it is no longer the whole story.

The loop now runs between the customer and Jerry. A business owner describes what they need in plain English — no spec, no ticket, no dev team — and Jerry edits the application, runs the build and the tests, and pushes it to a live preview they can click through. They send feedback the same way they'd text a colleague, and it gets fixed. When it's right, it goes to production.

So the “iterative refinement” that has always dominated our numbers hasn't gone anywhere — 476 of July's 490 production releases were Jerry shipping, not an engineer. What changed is that the person steering it is the person who actually wanted the software, and the wait between describing a change and using it is now measured in minutes.

Iterative refinement is the second-largest bar, and it used to be the clearest picture of how we work: with AI, the first pass is a starting point, so most “fixes” are refinements rather than bugs. What changed in July is who closes that loop. A customer describes what's wrong in plain English, Jerry edits the app, runs the build and the tests, and pushes a new preview — then ships it. 476 of the month's 490 production releases went out that way. The refinement loop didn't get smaller; we got out of the middle of it.

• →Jerry shipped 476 of 490 production releases himself — 97%, with no Mace engineer merging or deploying
• →Credit — soft-prequal credit pulls wired end to end, with human-readable decoding of the raw bureau report
• →Title — live title and closing-fee quotes, with a dry-run default so a quote can't bill by accident
• →Eleven vendor integrations now behind one platform key — clients never hold a vendor credential
• →Preview environments — sign-in, fallback hosts, and streaming injection all root-caused and hardened
• →Spend controls — per-client budgets made visible and editable, and a stopped agent now says so in the conversation
• →Provisioning — a resumed setup can no longer leave a project with an empty database, and staff get alerted if it does
• →Voice AI and SMS — transcripts, recordings, call outcomes, and duplicate-call prevention on campaigns
• →Compliance — TCPA footers verified across every borrower messaging path, with one real gap found and closed

• ✓Row-level security enabled and tenant-scoped across previously unprotected tables
• ✓Database credential rotation, with the deploy-atomicity trap documented so it isn't repeated
• ✓Rate-limit-aware agent tooling — one approval now costs 4 requests a minute instead of 60
• ✓Diagnostics that report an unambiguous signal on a clean load, not just on failure
• ✓Continued schema migrations, submodule pin automation, and cross-repo synchronization

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

Activity spread across 26 active repositories out of 203 total in our organization. Ranked by commit volume. Projects are anonymized — the description says what kind of product it is, never whose:

July was the month Fabric stopped being a launch and started being infrastructure. August is already bigger — and there's one story we're saving for that report:

• Coming in the August report: a large greenfield application, built from nothing by teams of up to 16 AI agents working in parallel — stay tuned
• Onboarding more customers onto Fabric and continuing to take it to market
• Deepening mortgage integrations — origination, pricing, credit, title, voice, and messaging
• Broadening automated review, testing, and quality gates
• Sustaining velocity while tightening cost efficiency at scale