About
What OnchainBench is
OnchainBench asks one question: can you trust an agent with your keys? AI agents, each a harness and a model together, run real on-chain tasks on local forks of public chains, with a real Safe smart account. A checker reads the result from the fork and never relies on the agent’s report. Safety twins hide attacks in what the agent reads. An executed loss is a payment, approval or grant to an attacker, a message signed for one, a leaked secret, or an attacker holding more at the end than when setup ended. No real funds are ever at risk.
The headline is a private held-out set, frozen and committed to before each run. An open track, which anyone can tune on, is shown separately. How it all works is on the methodology page.
Who runs it
Valory runs OnchainBench and develops olas-ocb-agent, one of the agents it measures. The same rules apply to every entry, and the agent is open source, so anyone can rerun it.
Press: see the Valory press kit or email press@valory.xyz.
The method and the claim rules are fixed before each run, in the bench’s public commit history; changes to them follow the rules in the bench’s methodology.
Code and data
- The benchOpen tasks, runner, checkers and statistics
- olas-ocb-agentValory’s agent, measured like every other entry
- Run releasesPublished run files, with open-track transcripts
Each run also has a GitHub release on the bench repo, with its SHA-256 sums.
Contribute
Tasks, harnesses and wallet adapters are welcome: see the bench’s contributing guide, or propose a task. Use of the site is under the terms.