Agent SkillOS

MontrealAI / skillos

the work-compounding layer for self-improving AI agents.

SkillOS now proves one concrete reference workflow gets cheaper, faster, and better as agents do the work and repeated corrections become tested skill releases.

See the proofView GitHub Actions
Target live sitehttps://montrealai.github.io/skillos/

GitHub Actions runs tests, generates data/wealth_proof.json, and publishes this dashboard automatically.

Pages statusTests status
Cheaper
Faster
Better

Workflow proof

Loading…Sales follow-up from call notes

Autonomous no-send proof

Shadow Pilot Proof

GitHub Actions now runs a no-send reference evaluation that shows SkillOS can turn call-note examples into traces, lessons, tested skill rules, and improved holdout performance.

No emails sent. No customers contacted. No private data. No API keys. Anyone can rerun the proof in GitHub Actions.

View visual proof Run on GitHub
+29.8 ptsquality gain
64.1%edit-time reduction
+80.0 ptsaccepted-rate lift
0%hallucination rate after learned skill
—

cost per job reduction

—

time per job reduction

—

quality gain

—

projected annual savings under demo assumptions

reference workflow proof

One reference workflow improves across tested skill releases.

The repository generates this proof during the GitHub Actions build. It runs a sales follow-up workflow, captures traces, converts repeated corrections into candidate skills, evaluates each candidate, releases approved versions, and checks that unit economics improve monotonically.

Before → After

Checking…

Plain-English result

Loading generated proof…

Skill versions: cheaper, faster, better

VersionQualityMinutes/jobCost/jobAcceptedLearned rules

How the proof works

Work becomes traces. Traces become skills. Skills improve future work.

  1. Work

    Agent drafts sales follow-up emails.

  2. Trace

    SkillOS records outputs, quality, time, cost, and human corrections.

  3. Learn

    Repeated corrections become lessons.

  4. Skill

    Each lesson becomes a bounded skill update.

  5. Test

    The candidate must improve quality while reducing time and cost.

  6. Release

    Only approved versions become current.

Approved releases

View generated proof JSON summary
Loading generated proof snapshot…

For GitHub web users

Upload it, click Pages, wait for green checks.

No terminal is required.

1

Create

Create a public repository named skillos under MontrealAI.

2

Upload

Upload everything inside UPLOAD_THE_CONTENTS_OF_THIS_FOLDER_TO_GITHUB.

3

Enable Pages

Go to Settings → Pages and choose GitHub Actions.

4

Open

Visit montrealai.github.io/skillos after the deploy check turns green.

What green means:

The Action verified the repo, ran unit tests, ran the reference workflow proof, generated data/wealth_proof.json, and deployed the dashboard.

Reference implementation

The proof is code, not just copy.

Agent Runtime

Runs jobs and creates traces.

Trace Store

Stores outputs, feedback, scores, tools, and skill usage.

Learning Engine

Finds reusable lessons.

Skill Trainer

Creates bounded skill edits.

reference workflow proof

Checks quality rises while time and cost fall.

Release Center

Publishes approved skill upgrades.

“Work stops being disposable and becomes capital.”