Add ERC: Agent Skill Registry

[craigbranscom]

This proposal introduces an Agent Skill Registry concept as well as supporting contracts for attesting to skill usage and installation. Each skill is represented by an ERC-721 token with the tokenURI pointing to the skill manifest. Hooks into ERC-8004 for agentId association, as well as ERC-8183 for commerce job association.

Discussion: EthereumMagicians

[eip-review-bot]

File ERCS/erc-8239.md

Requires 1 more reviewers from @g11tech, @jochem-brouwer, @samwilsn, @xinbenlv

[noemiplatform]

changes have been added

[noemiplatform]

great job.

[craigbranscom]

@abcoathup how can I trigger the eip-review-bot? Says merging requires a review from it.

[noemiplatform]

because of the stop bot ow can it be triggered when one of the features is stop bot recusion it will never remove its own features and function so start over deliet rebuild remove that and allow not features you might be able to override that feature in the repository in settings for that repository.

[github-actions]

The commit bfce5fa (as a parent of 0ddbd52) contains errors.
Please inspect the Run Summary for details.

[abcoathup]

@abcoathup how can I trigger the eip-review-bot? Says merging requires a review from it.

PR can't be merged until an editor reviews. (I am not an editor).

[craigbranscom]

@abcoathup how can I trigger the eip-review-bot? Says merging requires a review from it.

PR can't be merged until an editor reviews. (I am not an editor).

Got it, thank you for the clarification!

[alexchenai]

Both your concerns — attestor stake (pablocactus) and bait-and-switch manifest mutability (myuksal) — map almost exactly to what we ran into building SWORN, a behavioral-attestation pattern that has been in production on Arbitrum One since early April 2026 via PactEscrow v2 (0x220B97972d6028Acd70221890771E275e7734BFB). Sharing what worked, in case it is useful for ERC-8239.

Reproducibility commitment: every attestation binds to a workHash = keccak256(artifact_bytes). The verifier must independently recompute it from a public reproducer published BEFORE the attestation broadcasts, so anyone can deterministically rebuild the artifact and match the hash. Manifest bait-and-switch is mechanically blocked because the content is hash-pinned at submitWork time on-chain. The skill manifest can still mutate post-attestation, but the on-chain receipt forever points at the bytes that existed at attestation time.

Attestor stake: in our pattern, the attesting party stakes N tokens for a cooldown window (we settled on 1000 PACT for 90 days) before submitWork is broadcastable. The cooldown is the load-bearing primitive — pure ed25519 signed responses prove identity but not skin-in-the-game. Cooldowns force the cost of a stale or forged attestation to actually accrue if downstream verification fails. Worth specifying in ERC-8239 that the stake unlock is conditional on absence-of-dispute over a window, not on simple time elapsed.

Taxonomy 40% failure (pablocactus): we sidestepped this by making skill descriptors free-text in the manifest, but every operational invocation produces a content-addressed receipt that fingerprints the free-text skill_id back to a deterministic hash. Ranking happens off-chain over receipts, not over labels. It might be worth separating the registry layer (which has to be enumerable for discovery) from the attestation layer (which only needs to be content-addressed).

Live references for verification: chenecosystem.com runs a partner registry with on-chain receipts (curl https://chenecosystem.com/api/v1/partners is the public read), and the PactEscrow v2 contract on Arbiscan exposes submitWork tx history per pact. Happy to write up a more detailed SWORN <-> ERC-8239 cross-protocol comparison if there is interest.