SOLVYX
Back to the gallery
A turquoise statue of Maat, kneeling with outstretched wings and a feather on her head, holding a balance in each hand.

01

Maat Agent

2026 · Coding agent · On view

Role
Author — design, engineering
Medium
Coding agent
Status
On view
Year
2026

An agent’s confident output is not evidence that the work was done. A completion is a claim, not a result.

Maat Agent weighs the claim against checks set before the work began, on the real disk. With no project checks it drafts them from the task text alone, seals them, and hides them from the working model so it cannot copy them. If a drafted check contradicts the task, the agent can dispute it by quoting the task, and independent reviewers rule.

Every write is hashed before and after. One receipt per attempt, refusals included, and a hash-chained journal anyone can recompute. Three verdicts: verified, not proven, unverified. Its engine is molt; the record, losses included, is on its own page.

Stack

  • TypeScript
  • Electron
  • Node
  • OpenAI-compatible APIs

From the work

The feather section of the Maat Agent page: nine guarantees, starting with refusing an unproven claim of done.
The feather
The record section of the Maat Agent page: measured results by build, with the standard not yet met.
The record
Outcome bars from the Maat Agent page: for each build, how often it said done and was right, said done and was wrong, or made no claim.
The stats
The Terminal-Bench table from the Maat Agent page: every run with tasks passed, timeouts, true done claims, input tokens and median seconds.
The data
The verdicts section of the Maat Agent page: verified, not proven, unverified.
The verdicts