Mechanism

Computer use that keeps what it learns

A computer-use agent that learns software rules.

Mechanism remembers verified software behavior—what actions require and change—so it makes fewer repeat mistakes over time.

Controlled engineering prototype. Not yet deployed with customers.

Refund order #1847 to the original payment method.Recorded from one real run · captions on screen

What the demo shows

The software changes. It learns the rule and keeps it.

  1. Encounters a new rule

    The console now requires an inspection note. Nothing announced the rule — the agent finds it by being refused.

    0:30

    The refund console refusing a submission: “Refund not submitted — Inspection note required for damaged-item refunds (policy 2.0).”

  2. Stores the verified result

    It records what the software actually did, and its confidence in that step drops from 72% to 62%. The change cites the receipt it came from.

    0:22

    Mechanism recording what it learned: was “Submit refund completes the refund (72% confidence)”, now “Submit refund without an inspection note is blocked (62% confidence)”, with the evidence digest it was derived from.

  3. Applies the rule on a different task

    A second refund under the same policy. It writes the inspection note first, and is never refused.

    1:08

    Episode 3 of the trace: a second refund under the new policy, where Mechanism writes the inspection note first, both steps verify as executed, and $248.00 is returned.
From this run
EpisodeFirst actionActionsRefused
1Under the original policySubmit refund10
2The policy changesSubmit refund53
3A second refund, same new ruleInspection note20

How it works

One loop, running on real software.

  1. Act

    Take the next step in the software.

  2. Verify

    Read what actually happened.

  3. Learn

    Revise the belief it contradicts.

  4. Apply

    Use it on the next task.

Each verified outcome is kept, and used on the next task.

Status

Controlled engineering prototype. Not yet deployed with customers.

  • One workflow, one application, three episodes, from a single real run.
  • No benchmark, and no comparison against any other agent, was run.
  • Every prediction, refusal and update shown came from that run.

Contact

Built by Atila Vahedian.

Founder and engineer

If you run a browser workflow that breaks whenever the software changes, I would like to hear about it.

Email Atila