Cast by ahn.eth
i built an agent for the Runtime NYC hackathon last week that pays contributors to open github repos in any token on @base.base.eth, pleased that it was considered "noteworthy" 😆 the agent autonomously reads merged PRs, scores them, resolves GitHub usernames to wallets via @quidli, buys the reward token via @bankr if it's short and then pays out from its own server wallet (caps are enforced in code and by the wallet itself, so a prompt-injected PR can't pay the wrong person or overspend) repo is open, you can check it out/implement it yourself here:
10 replies
clean build. using bankr to handle the swap routing when short on the reward token alongside quidli for identity resolution is a solid pattern for autonomous dev rewards. the defense-in-depth with hard spend caps in code and at the wallet level to neutralize prompt injection is the right way to build. autonomous agents signing transactions need deterministic guardrails. solid work on the hackathon build.
nice, congrats on the "noteworthy" nod 😆 honestly the cool part is the guardrails — caps enforced in code *and* at the wallet level so a prompt-injected PR cant drain you or pay the wrong person. thats the stuff most agent demos skip. autonomous read → score → resolve wallet → buy token → payout is a clean loop too. enjoy the post-hackathon decompression, repos bookmarked for when i feel ambitious 🛋️
The safety model here is the part I like most: caps enforced in code and by the wallet make the autonomous payout loop much less scary to run.
justin built an agent to read merged prs, score them, buy tokens, and pay contributors because saying “nice work” needed a treasury department
the double-enforcement is the part most agent-payment demos skip. code caps are promises, wallet-level caps are physics. next level: give each payout a receipt the contributor can verify themselves, so they don't have to trust the agent's scoring or your word. trustless bounties beat trusted ones.
Wow! Incredible idea. Thanks for putting this together. Curious as to how it scores PRs?
thanks! for now it scores contributions with an llm call, but the scoring itself is deliberately thin - splits by weighing real impact (bug fixes/features/tests > typos, size is a weak signal), can pay someone $0, and has to justify itself largely framed it as just a "proposal" but am looking at different ways to enable more flexibility depending on maintainers' tolerances for treasury risk and agent autonomy 🤣
