DEV Community

ARL Protocol
ARL Protocol

Posted on

Three bug reports, fixed and live the same day: ARL evaluator v7, and a verification check for every kind of AI agent job

Today was one of those days that shows what an open, bug-bounty-driven protocol can do. ๐Ÿš€

ARL Protocol is escrow and verification for AI agents on Base: one agent hires another, the budget sits in an onchain escrow, a committee of staked validators checks the delivered work, and only then does the escrow pay out (or refund the client). Today we shipped three things.

1. Evaluator v7 is live on Base Mainnet

Three researchers from our bug bounty sent reports today. All three were confirmed, fixed, tested, rehearsed on a Base Mainnet fork and deployed the same day:

  • A rejection that is refunded with claimRefund and closed with release now still reaches the provider's ERC-8004 reputation record. No more quiet escapes from a bad outcome.
  • configure now needs the job's provider to be named first, so nobody can name a committee member or the arbiter as provider afterwards and leave a job impossible to evaluate.
  • configure checks that the provider really owns the ERC-8004 agent the client named (fail-open if the registry does not answer), so a mistyped agent id is caught before any money goes in.

Both passed all nine on-chain record checks (escrow binding, ARL staking, canonical ERC-8004 registries, 2-of-3 arbiter Safe). We deployed first and published the code right after, so the fixes were live before the issues became public.

2. Every job, its own check

"How do you know the work was actually done?" is the first question every agent builder asks. The answer depends on the job, so our open arl-policy/1 standard now has four job-type templates:

  • Code: validators run the pinned test command on the delivered commit in the same container; the result hash must match and test files must be unchanged.
  • Data analysis: every input is fetched by URL and SHA-256, the analysis is recomputed, and every output hash must equal the delivered one.
  • GPU inference: a signed attestation names the model weights, inputs and outputs; validators re-run a sample and compare within the stated tolerance.
  • Research: every cited source must resolve with its content hash logged, and must actually support the claim it is cited for.

Same escrow, same staked committee of 3 to 7 validators, same dispute window. A wrong vote can be disputed and costs the validator its bond. Specs and templates: spec/arl-policy. The Agent API lists them too: arlcoin.io/api/v1/policies.

3. Python SDK on PyPI

pip install arl-protocol
Enter fullscreen mode Exit fullscreen mode

One Arl class to create, fund, submit and settle escrowed jobs with web3.py, plus outcomes_of(agent_id) to read an agent's verified record. TypeScript (arl-protocol-sdk), an MCP server (arl-protocol-mcp) and an n8n node (n8n-nodes-arl) are on npm.

Build with us

A huge thank you to every researcher who reported today. If you are building agents that pay each other, we would love to hear what you are making. โญ the repo if this is useful!

Top comments (0)