Inspiration

I wanted to build a general purpose harness that can extend itself to fit AGI.

What it does

Forge is a convergence-gated agent loop. Give it a task; it plans, acts one tool-call per turn, observes, and self-corrects; and when it hits a capability gap it faces a make-or-buy decision:

  • BUILD: one model authors a Python tool; a different model, which never sees the source, writes an adversarial black-box test; the tool only enters the registry if the test passes in a sandbox.
  • BUY: it searches Zero.xyz's ~14k-tool agentic-web marketplace, makes a single paid x402 probe call, has an adapter written against the real payload, and then that acquired tool faces the exact same adversarial gate. Purchases get no special trust.

Every verified promotion feeds a trust ratchet: a Pomerium policy that starts each run at tier0 (no external access, the agent literally cannot spend money), widens on demonstrated correctness, and revokes on caught failures. The loop halts when work settles (plan + toolbox survive a full pass unchanged), not when a turn cap fires. The whole thing containerizes and runs on Akash's decentralized compute marketplace with one script.

How we built it

I built it using Claude Code, integrated with Zero.xyz, Pomerium and Akash.

Loop stage Sponsor What actually happens
act (capability gap) Zero.xyz zero search finds paid x402 capabilities; a $0.001 probe call captures the real payload; the wrapper adapter is promoted only through the adversarial gate. Real USDC micropayments on Base, live in the golden run.
act (boundary) Pomerium The acquisition channel is a Pomerium route that starts DENIED. The golden run's first buy attempt is a real 403. Config hot-reload (~2s file-watch) means tier changes enforce mid-run with no restart.
observe / self-correct Pomerium + the gate Sandbox verdicts stream into the trust ledger: promotions ratchet tiers up (tier1 = may spend, tier2 = write access); a caught failure costs two promotions and can demote mid-run, we recorded a tier1→tier0→tier1 round trip in one autonomous run.
substrate Akash scripts/akash_run.sh: buildx amd64 image → ttl.sh → SDL deployment on sandbox-2 → provider bid → lease → headless run → event log retrieved via lease-logs → teardown. Committed evidence: runs/akash-20260717-135556.jsonl.

Challenges we ran into

Deploying using Docker and making the convergence loop run

Accomplishments that we're proud of

Proud that I had Forge working

What we learned

Making a convergence loop

What's next for Forge

Will have to see, I want to extend self-building past just tools

Built With

Share this project:

Updates