AI agents often report success because a tool call completed, even when the actual outcome is wrong or incomplete. DidWork independently verifies agent work against the expected result, gives agents clear pass/fail gates, and can trigger repair when verification fails. It works through MCP and an SDK, supports multiple providers, and produces shareable verification receipts so developers can see what was checked and whether the work was actually done.

I built DidWork because I kept running into the same problem while using agents inside Pulltrader: the tool call succeeded, the agent said the work was done, and the actual outcome was still wrong. That happened enough times that I stopped treating it like an edge case and started researching it. I ended up documenting 38 failure claims across 8 providers, with the same pattern showing up over and over: agents are usually very good at knowing whether an action ran, but much worse at knowing whether the intended result actually happened. DidWork started as an internal tool to solve that. The idea is simple: give an agent an independent verification step before it moves on. Check the outcome, return a clear pass or fail, and give the agent a chance to repair the work when verification fails. I initially built it for myself, then realized this problem is going to become more important as we give agents more autonomy. Would love to hear how other people are handling this today, especially if you’ve had an agent confidently tell you something was finished when it definitely wasn’t.

Thanks to everyone who emailed us over the last day about Codex permission issues. We found cases where Codex’s approval reviewer blocked verification because it didn’t recognize authorization to send private GitHub repository metadata to DidWork. We’ve shipped updated Codex/Claude and Cursor plugins with explicit, repository-scoped consent during onboarding, saved for future tasks. Testing also caught a credential-forwarding bug in the Codex plugin, which is now fixed. Local Codex tests passed for both requested and automatic verification across fresh tasks without repeating consent. We’ll keep testing broader onboarding scenarios.

DidWork

Verify agent outcomes, not just tool calls.

Verify agent outcomes, not just tool calls.