Smarter gates, leaner agents, faster everything: this week in the Agency Orchestrator

Week of October 4–11, 2026

An AI agency that ships fast is easy to build. One that ships fast and stays honest is harder. This week we worked on both: deploys now wait for the reviews that should come first, specialist agents carry far less context, and the hooks behind every action got a lot faster.

🚦 No review, no release

Agents are confident, sometimes more confident than they should be. So we stopped relying on them to remember the rules and wrote the rules into the tooling.

The orchestrator now has a pipeline gate. Nothing deploys and nothing goes public until the architecture and security reviews are actually on record. It covers the major platforms: Vercel, Expo/EAS, Firebase, Google Cloud, Cloudflare, Fly and Netlify. There's no hidden override. Skipping a step needs a written reason, and that reason is logged.

Alongside it come new provenance tools. One scans a project for anything that shouldn't go public, like local file paths, personal emails or unverified claims. The other checks what an earlier AI session actually read or ran, straight from the session record rather than from notes.

The idea: "verified" should point to evidence you can check. It shouldn't be a label an agent writes for itself.

🪶 Specialist agents, now with up to 78% less baggage

Every specialist agent (designer, SEO, PM and the rest) used to load the orchestrator's full instruction set, mostly rules meant for someone else. Now each one loads a small shared rules skill with only what its job needs.

In the first test, the diagram agent's starting context went from 26.6k tokens to 5.9k (−78%), and its output still rendered perfectly. Eight more specialists have switched over, and each now starts with about 5–6k tokens.

Less context means lower cost, more room to work, and agents that stay on their own task.

⚡ Hooks up to 88% faster

The orchestrator runs safety checks behind every action. They're important, but on Windows they were slow. We rewrote them to stop starting extra processes on every call:

• Billing check: 1.6 s → 0.2 s (−88%)
• Per-action checks: about 35–40% faster
• Total overhead per prompt: about 1.6 s → 0.94 s

The checks now run about as fast as Python itself can start, so they're no longer slowing anything down.

🔄 Updates across projects, done right

One orchestrator keeps 14 projects up to date. This week we fixed every way that sync could fail while reporting success:

• ✅ It won't build on outdated code. If a project is behind, the sync stops.
• ✅ It reports push status correctly in the project registry.
• ✅ Exclusions actually exclude, and a typo in a project name triggers a warning.
• ✅ New files ship with the update instead of being left behind uncommitted.
• ✅ It warns about a machine-wide package-manager setting that had broken hundreds of package links across projects.
• ✅ A new read-only report finds work left in old stashes. Its first run recovered scripts that existed nowhere else.

🧪 Tested so the tests can fail

We don't stop at "the tests pass." For each fix, we turned the fix off and made sure the new tests failed. That proves they catch the real bug. At the end of the week the suite stood at 380+ automated tests, plus the repo's lint checks.

The bottom line: the agency is faster, cheaper to run, and has to show its evidence before it releases anything.