Repository navigation
Show and tell: a check_work function tool — my ADK agent can't call its work "done" until it's been reviewed #8000
Fikpeze123
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi — I'm Frank, founder of Pactlyst.ai (built with AI help, openly). Sharing a small tool I give my ADK agents, because the failure it fixes is one I suspect is common here.
My research agent produces a short customer migration plan: sources in, a tidy document out. The plan read beautifully. It also promised a data-residency term we'd never agreed to — fluent, confident, and wrong in a way nobody caught, because the last step in the agent was "return the document." Now the last step is a tool call:
The check reads the artifact against playbooks for its kind — documents, code, and software work all go through this one door — and grades it against the ask it was given, line by line. The gate is the part that matters in production: Low risk proceeds; Needs review waits for a named human to acknowledge the findings; High risk waits for a named approver who isn't the requester. Either way the call returns a signed PactCheck Receipt™ and a public receipt URL you can attach to the session's output, so the artifact travels with its proof. The artifact text itself is never stored — hash and counts only.
This is not an official ADK integration — it's a function tool wrapping one HTTP call, with a free key. ADK agents that prefer MCP can reach the same check as a
check_worktool on Pactlyst's MCP server.I'm opening a few design-partner spots for agent builders — three months of Pro free. Docs at pactlyst.ai/docs. And a real question for this board: do you gate your agents' final artifacts today, and if so, with what?
— Frank
All reactions