Pixl Works.

Tinkerflock

From idea to deployment.

Our own agentic coding system. A swarm of AI agents plans a task, builds it, tests it, reviews it, ships it and checks the result. If that check fails, it undoes the release by itself. We build with it every day.

970 tasks shipped · 16 projects · 2 machines · as of 17 November 2026, read from the running engine

Try to break it

Let a step fail or pull the plug, and watch what the swarm does.

A heart is one way back to build. When they are gone, a person takes over.

  1. spec_check

    Check the brief

  2. build

    Build

  3. gate

    Gate

  4. security

    Security

  5. quality

    Quality

  6. review

    Review

  7. regate

    Second gate

  8. merge

    Merge

  9. deploy

    Deploy

  10. verify

    Verify

  11. deployed

    Done

  • rollback

    Rollback

    The swarm restores the version from before the release.

  • escalated

    With us

    What the swarm cannot solve lands with a person.

FLK-204

Check the brief. Is the task clear enough to build? If not, it goes back to a person.

The task runs by itself until you press a button. After that it is your call. This is the real workflow of one of our projects, with its real limits. The task FLK-204 is made up.

All steps as a list
  1. Check the brief. Is the task clear enough to build? If not, it goes back to a person.
  2. Build. A coding agent writes code and tests, in a sandbox without open network access.
  3. Gate. Lint, strict type check, tests. At least 80% of the changed lines must be tested. On red back to build, at most 3 times.
  4. Security. Credentials in the code, risky patterns, vulnerable dependencies. On red back to build, at most 2 times.
  5. Quality. Limits on complexity: code nobody can read does not get through. On red back to build, at most 2 times.
  6. Review. A second model reads the change, by default from a different vendor than the builder. On red back to build, at most 3 times.
  7. Second gate. The same checks once more, right before the merge. On red back to build, at most 2 times.
  8. Merge. The change lands on the main line. If that has moved meanwhile, the gate runs again.
  9. Deploy. The new version goes live.
  10. Verify. Does it really run? If not, the swarm restores the previous version by itself.
  11. Done. Live and verified.

Coding agents are fast. Reliable they are not.

Six ways a coding agent causes trouble. And what stands against each in the workflow.

The Bluffer

Reports "done" while the tests are red.

Gates, not trust

The checks run apart from the agent that built. It cannot skip them or talk its way past them.

The Flood

Writes faster than anyone can review.

A second model reads

Whoever builds does not review itself. The review is done by another model, by default from a different vendor.

The Crash

Dies halfway, and the work is gone.

Save points

The state of every step is saved. If the system goes down mid-build, it picks up right there.

The Runaway

Reaches into the network where it should not.

Sandbox

Agents build in isolation. They reach the network only through an allowlist.

The Silent Bug

All green, and live it still does not run.

Rollback on its own

After shipping, the swarm checks that it really runs. If not, it restores the previous version.

The Endless Loop

Tries the same thing again and again.

Counted tries

Every way back is counted. What the swarm cannot solve escalates: first to a supervisor agent, then to us.

The agent may be fast. The workflow decides when something is done.

The numbers, awkward ones included

970
tasks are live and verified.
878
of them went into Tinkerflock itself. It keeps building itself, under the same rules.
398
were cancelled, for example because a re-cut task replaced them.
4
failed. That is in the log too.

As of 17 November 2026.

What it cannot do yet

STILL LOCKED

Interface from a design mockup

The code is right, the tests are green, and it still does not look like the design.

We are working on that right now. Until then we check every layout ourselves.

59 seconds of trailer

The trailer is from 17 November 2026 and counts up to 970 shipped tasks. The film is code too: an agent wrote the picture and the score.

And what is in it for you?

Tinkerflock is not for sale. It is our workshop.

But its rules are the ones we build agents for your business by: fixed steps, checks an agent cannot skip, a person at the points that matter, and a log of everything.

The homepage of this website, by the way, was built by Tinkerflock's predecessor.

Free first call