Blog · Factory

Jul 23, 2026 The Factory earns its keep: ready to daily-drive A full-fleet test pass, a live plan-build-verify run on real infrastructure, and the features that got us here: baked toolchains, cached code graphs, per-spec isolation, and verdicts you can trust.

Jul 19, 2026 See it run: a visual walkthrough of one autonomous build The same live run, in pictures and one continuous video. A plain issue in, a tested pull request out, across four portals and one cockpit.

Jul 19, 2026 A factory that builds software: one live run, end to end A plain GitHub issue went in and a tested pull request came out, with no human in the loop. Here is the run, the four stages, the real verdict numbers, and the moment the tests refused to lie.

Jul 13, 2026 Measuring ourselves against the field, then fixing what we found A July 2026 cycle report: the first external benchmark number, the honest failure it exposed, and the cost-aware routing we shipped and proved as a result.

Jul 7, 2026 Four portals, one product: the Factory becomes a single shell A unified cross-portal shell, one sign-on, fleet-wide search, and the whole factory moved to the latest models — with a one-page showcase for each service, ready to download.

Jun 27, 2026 One plan, two clouds: the Factory ships a live app to GCP and Azure through its own pipeline PFactory plans it, AIFactory builds it, a GitHub Actions runner deploys it, TFactory plays the game and checks the score — and we have the screencast to prove it

Jun 26, 2026 The test factory tests the factory: TFactory now drives the Factory portals through MFA Real Keycloak logins, every menu and dialog exercised, screenshots and verdicts surfaced where humans already look

Jun 23, 2026 The last node-pin: making a packed build node-agnostic Packing the workspace to object storage freed /work — but the build still nailed itself to one node through the Nix store. Here is how we cut the last wire, and the honest line on what is proven.

Jun 21, 2026 Program update: week of June 15–21, 2026 SpecKit-informed interop and the concurrency core shipped; Job-native defaults are converging — the honest fleet-wide status across PFactory, AIFactory, TFactory and CFactory

Jun 20, 2026 Earning the right to run unattended: the hardening behind the AWS demo A real OS sandbox, seven autonomous-completion fixes, and more verification surface — the trust work that let one brief reach a live EKS cluster

Jun 19, 2026 Plan to production on AWS: a 3-tier app through PARR, deployed and torn down One brief becomes a Fastify + Postgres + Redis service running on a real EKS cluster — including the parts the generator got wrong

Jun 18, 2026 Planning that reads your code How PFactory stopped guessing — reconnoitring an existing repo to plan real changes, and proving a Python-to-Rust rewrite behaves the same (RFC-0010)

Jun 18, 2026 Demoing the Factory: one MFA-gated change, from plan to proof How we showcase the whole PARR pipeline on a single unit of work that hides behind two-factor auth — the planning logic, the 2FA-authenticated test, and the evidence in the cockpit

Jun 17, 2026 Seeing the proof: test evidence with screenshots and recordings in the cockpit How the Factory went from 'the tests passed' to 'here is the page, here is the click, watch the recording' — visible in both TFactory and CFactory

Jun 14, 2026 Bring your own Ollama: running Factory builds on your own models through a Cloudflare Tunnel A remote user wants the Factory to code with the 70B model sitting on their own workstation — not our hosted providers, not an API bill. The honest answer is that you cannot connect a laptop to a cloud agent directly, and WebMCP is the wrong tool. But there is a real, scriptable pattern: expose your local Ollama through an outbound Cloudflare Tunnel and hand the Factory the HTTPS URL. Here is the full automation, why it fits our architecture with no infra change, and an even-handed case for when you should and should not do it.

Jun 13, 2026 Seeing the Factory think: per-worker, per-provider observability — and the hardening wave that made it safe to ship A 16-agent deep audit set off a security and CI hardening wave across the suite, then a super-brainstorm produced the missing leg: observability. The key finding was codebase-first — AIFactory already ran parallel multi-provider coding workers; we just couldn't see them. So we built additively: per-worker cost and OTel metrics, live worker events, a CFactory cockpit with a ticking per-task cost stamp, a soft budget alert, and OpenObserve bundled as the OTLP backend. You can finally answer what a build cost, on which model, and where the time went — live, per task.

Jun 12, 2026 Field report: we put the Factory on a benchmark — and it found six bugs in our own harness first Phase 1 of the make-it-great plan was 'prove it.' So we pointed our PARR benchmark at the live deployed fleet for the first time. The build leg works — a real autonomous service in ~23 minutes on the cluster — but first contact surfaced six cross-service seam bugs the per-repo unit tests never could. That's exactly what the benchmark is for. Here's the honest scorecard, and the verify-leg gap we closed the same day.

Jun 12, 2026 The honest scorecard: we reviewed all four factories against the market A full code review of PFactory, AIFactory, TFactory and CFactory against our goals and against Devin, Factory.ai, OpenHands and the cloud agents. The verdict: the spine is closed and we're ahead on the axis that matters — but we have zero published proof, zero design partners, and one security item we refuse to hand-wave. Here's the plan.

Jun 11, 2026 Proof: the Factory logs into a live web app, finds the bug, and screenshots it We deployed an authenticated web form to real AWS, had a TFactory-style browser test log in as a test user, exercise the UI, catch a deliberate validation fault, and record screenshots as evidence — then tore it all down. If it can do this, it can test and record almost anything.

Jun 11, 2026 Watch it deploy: from a CLI to a live, tested API on real AWS — and torn down We closed the last gap in the PARR verify leg. This screencast shows the Factory taking services from a command line, deploying them to real AWS App Runner with deterministic Terraform, proving the live API works, running the acceptance tests against the live endpoints, and tearing everything down — cost-guarded, scoped, repeatable.

Jun 11, 2026 Field report: the first end-to-end Gemini PARR run We pushed a real job — an EKS Terraform module plus three FastAPI microservices — through the whole factory chain on Gemini: plan, govern, sign, build, hand off, verify. Here's what worked, what broke (and got fixed live), the one gap that's left, and where we go next. An honest field report, not a victory lap.

Jun 11, 2026 The guards behind the pipeline: how Factory turns agents into a process A new reference page maps every step and decision from handover to merge — and the fifteen guards that make AI-built code something an enterprise can actually adopt. Plus four real adoption scenarios, from the solo builder to the regulated enterprise.

Jun 9, 2026 Ship day: the Hermes learnings land — and a lot more In a single day the four-product suite shipped agent-loop hardening, CI-parity verification, per-user API keys, an Ollama Cloud provider in every product, and a cockpit that finally tells the truth. A grounded changelog — with what's on main and what's still on dev.

Jun 9, 2026 What the Hermes agent taught us We read Nous Research's open-source agent end to end. Here's what we're stealing — and what reading other people's code does for a project like ours.

Jun 7, 2026 Keycloak, GitHub SSO, and the 401 that lied to our faces How we gave the Factory suite a front door — and the three-headed bug that kept slamming it shut

Jun 5, 2026 From Backstage to verified code: building an API gateway with the Factory PARR pipeline One governed unit of work flowing through PFactory → AIFactory → TFactory, watched in the CFactory cockpit

Jun 4, 2026 One factory, four products: why we split the SDLC into a chain PFactory plans, AIFactory builds, TFactory verifies, CFactory watches. Here's why that's a family, not a monolith.