Yesterday I posted here about not trusting a 200 from a tool call. A few people in the comments pushed the idea further than I had, so this is where I landed. When an agent says "done", it can mean any of these: Intended: it decided to do the thing Attempted: it called the tool Accepted: the API returned 200 Applied: the real system shows the change (Sent folder, the row, the calendar event) Correct: the change matches what the person asked for (right contact, right amount, right thread) Effective: the outcome happened (delivered, paid, they showed up) Most agents report 3 and call it 6. The read-back from my last post gets you to 4. The part I was missing was 5, and it came from that thread. A read-back compares the system against the agent's plan. If the agent picked the wrong contact from the CRM, the email goes to the wrong person, the read-back matches the plan, and it gets marked verified. Everything passes and it's still wrong. So for anything that sends, pays or deletes I now do two checks: system vs plan, then result vs the original request. 6 is mostly outside your control. You can't make someone read the message. But you can stop collapsing all six into one word. WhatsApp has done this forever with its ticks: sent, delivered, read. It never claims more than it knows. So the status my agents show now is "accepted by the API", "in Sent", "sent to the contact you named", or "delivered" once a webhook confirms it. Never just "done". Two questions for people running agents that take real actions: Which step does yours actually stop at today? For step 5, what do you compare against? The raw user message, a structured intent you extract up front, something else?
A practitioner shares a pattern where an AI agent's tool call returns success (200/id) but the intended action never actually completed (e.g., email stuck in drafts). They propose read-back verification after every write, labeling actions as verified, acknowledged, or contradicted, and ask how others handle agent action verification.
The article explores an early concept called agentuptime, which addresses verifying AI agent actions by independently checking outcomes to ensure that an agent's completion claim matches the actual state of external systems.