The Receipt That
Said No.

A second agent joined a governed chain that a first agent had already opened, inside a real IBM watsonx Orchestrate deployment — and the platform let it say no. The denial became a receipt, hash-linked to the one before it, in the same chain anyone can check.

The problem nobody's stack solves

Every vendor can tell you what its own agent did. Almost none can prove what happened once a second agent joins in — a collaborator spun up to handle a sub-task, acting under its own identity, inside the same logical run. That is exactly where multi-agent designs get restricted to a single approving agent in audited environments: nobody can sign off on delegation they cannot verify afterward.

We ran that delegation for real, gave the second agent a genuine reason to be denied, and made every step independently checkable.

The chain

What actually happened

1. A primary agent, running inside IBM watsonx Orchestrate, registered itself and asked whether an action was allowed. Real production policy path, real hash-bound receipt. Verify the root receipt →
2. The same agent asked again about a different action. Allowed again, chained off the first receipt's hash — not a fresh, disconnected record. Verify the second hop →
3. A genuinely separate collaborator agent — its own credentials, registered as a child of the primary — joined the same chain and asked about an action the organization's policy denies. It got told no, and the denial is a receipt too, chained off the hop before it. Verify the full chain →

Scale honesty

This was an engineering verification run, not customer production traffic — a real multi-agent deployment, exercised deliberately to prove the mechanism holds under a real denial. What it demonstrates is narrow and real: a delegated agent's decision, including a denial, carries the same independently checkable evidence as the agent that spawned it. The mechanism scales; this run doesn't claim the scale.

Why it matters

For teams building multi-agent systems, this changes what's possible without changing what's required. A delegated agent's decision — including a real denial — now carries the same independent, checkable evidence as the agent that spawned it. Teams that would otherwise keep multi-agent designs narrow because a compliance reviewer couldn't sign off on unaudited delegation can build the collaborative pattern first and get the audit trail with it, not as a separate project bolted on after.

PromptKing decides. Your stack executes. The evidence makes the recorded decision independently checkable — across agents, across a real denial, in the same graph.