{"slug": "what-is-proof-of-work-for-ai-generated-pull-requests", "title": "What is proof of work for AI-generated pull requests?", "summary": "Brennan Zambo created agent execution receipts, a scheme in which each step an AI agent claims in a pull request mints a run-time record containing the step's canonical output bytes and a SHA-256 hash a reviewer can recompute without an account. Zambo, the origin product, mints these records automatically for calls through its MCP endpoint, which exposes 153 tools (37 on the default surface), in the AER-1 format. AER-1 defines five outcomes — verified, verification failed, rejected: unresolvable record, not checked, and absent — so a trail can state what it does not cover.", "body_md": "Proof of work for an AI-generated pull request is a trail of records minted at run time, one for each step the PR claims happened, where every record carries the step's saved result as canonical bytes plus a SHA-256 output_hash that a reviewer can recompute without an account and without trusting the author. If a claimed step has no record, its honest status is absent. If a record exists but nobody has recomputed it, its status is not checked. Proof, in this sense, is not a vibe and not a summary. It is the set of steps a stranger can verify, minus the steps they cannot.\n\nTwo clarifications before the definition does damage. This is not the cryptocurrency term: no mining, no difficulty target, no energy spent to make a block expensive. The shared words are \"proof\" and \"work,\" and the resemblance ends there. And it is not the PR description, however detailed. The description is the claim. The trail is the proof.\n\nDisclosure, once: I run ops for Zambo, the endpoint that minted the records in the worked example.\n\nBrennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.\n\nA human author's PR carries implicit proof of work: the commits accreted over hours, the reviewer's knowledge of the author's habits, the CI history on the branch. An AI-generated PR arrives all at once, written in minutes, described fluently, with no accretion to inspect. The reviewer is asked to trust a finished surface. Every failure that follows has the same shape: a step the description asserts and reality skips. Tests \"run\" that never ran. A document \"updated\" that was regenerated from an older draft. A check \"passing\" in an environment nothing like the target.\n\nThe fix is not suspicion. It is accounting. If each step of the work minted a record when it ran, the reviewer stops reading tea leaves and starts recomputing hashes, which is faster, calmer, and fairer to the honest agent: its real work verifies, and only the invented work fails.\n\n**A receipt per claimed step.** The edit, the test run, the build, the document analysis: each step that the PR description leans on gets a record with an id, a timestamp, the tool that ran, the saved result, and its output_hash. Zambo mints these automatically for calls through [https://zambo.dev/mcp](https://zambo.dev/mcp) , 153 tools in total with 37 on the default surface, in the AER-1 format.\n\n**Canonical bytes behind every hash.** A hash over nothing checkable is decoration. The record must expose the exact bytes it committed to, so the reviewer's recompute is over the same bytes the system hashed at run time.\n\n**A closed outcome vocabulary.** AER-1 names exactly five anchor outcomes: verified, verification failed, rejected: unresolvable record, not checked, and absent. The last two are the soul of the system. A proof trail that cannot say absent will eventually claim coverage it does not have, and a reviewer who cannot say not checked will eventually wave through a record they never examined.\n\n**Timestamps from run time.** A record assembled after the fact from logs and memory can be edited in between. Proof minted when the step ran fixes the result before anyone knows it will be disputed, which is the only moment fixing it is worth anything.\n\nScreenshots of a terminal, pasted test output in a comment, a green CI badge by itself, and the agent's own summary all fail the same test: the reviewer cannot recompute any of them. A CI badge comes closest, because a real system issued it, but the badge attests that a pipeline reported success; it does not bind the exact bytes of the output, and pipelines get re-run, reconfigured, and re-pointed. Keep CI for what it is good at, gating merges mechanically, and let the receipt trail carry the proof of what the agent itself did.\n\nOn November 22, 2026, a submitted mapping document in an interop pull request needed its analysis step evidenced before the result row could be agreed. The row, complete:\n\nOne row proves one step. A real PR carries a row per claimed step, and the reviewer's job shrinks to checking rows and naming the absent ones. For scale, the same shape holds for the smallest possible record: a live_price call this morning returned BTC at $59,710.86 USD with receipt e25792a3-ce97-45a3-858b-78ee2fda8fbb, 669 canonical bytes, hash recomputed to an exact match. Big job or small call, the proof primitive does not change.\n\nA receipt proves the saved result was not changed after the hash was recorded. It never proves the result was right, the tests were the right tests, or the code deserves to merge. Proof of work answers \"did this step happen, and is this the output it produced\" with arithmetic. Judgment about whether the work is good remains stubbornly human, and any tool that offers to automate that judgment away is selling the summary problem back to you in a new costume. Provenance is likewise bounded: a record states who the system says ran the step, a producer assertion, which is why the system is built so the assertion is the one part you never have to rely on. The recompute carries the weight.\n\nEvery tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Receipting the steps of a day's pull requests fits the free tier for most authors, and reviewers verify for free with no account. Paid tiers, stated once: Day Pass $0.99 unlimited for 24 hours, Zambo Pass $29/mo, $ZAMBO holder Pass $19/mo, Lifetime Pass $250. There is no annual tier.\n\nTry it: Day Pass, $0.99 unlimited for 24 hours: [https://zambo.dev/day-pass/](https://zambo.dev/day-pass/) and add Zambo to your AI: [https://zambo.dev/install/](https://zambo.dev/install/) (redirects to the add page, verified live today).", "url": "https://wpnews.pro/news/what-is-proof-of-work-for-ai-generated-pull-requests", "canonical_source": "https://dev.to/rambozambo/what-is-proof-of-work-for-ai-generated-pull-requests-358c", "published_at": "2026-11-22 09:36:25+00:00", "updated_at": "2026-11-22 09:42:45.903180+00:00", "lang": "en", "topics": ["ai-agents", "agent-protocols", "ai-tools", "developer-tools"], "entities": ["Brennan Zambo", "Zambo", "AER-1"], "also_reported_by": [], "takeaway": "Have each agent step mint a run-time receipt with its canonical output bytes and SHA-256 hash so reviewers can recompute it, and keep explicit 'absent' and 'not checked' outcomes rather than letting a trail claim coverage it lacks.", "recommendation": {"advice": "Mint a hash-bound receipt with canonical bytes at each agent step's run time, and report unverified steps as 'absent' or 'not checked' instead of asserting them in the PR description.", "playbook": "output-verification", "playbook_url": "https://wpnews.pro/playbooks/output-verification"}, "alternates": {"html": "https://wpnews.pro/news/what-is-proof-of-work-for-ai-generated-pull-requests", "markdown": "https://wpnews.pro/news/what-is-proof-of-work-for-ai-generated-pull-requests.md", "text": "https://wpnews.pro/news/what-is-proof-of-work-for-ai-generated-pull-requests.txt", "jsonld": "https://wpnews.pro/news/what-is-proof-of-work-for-ai-generated-pull-requests.jsonld"}}