The agent did the work.
Someone will ask what it did.
When an agent moves money or changes a record, the counsel, the underwriter, the auditor, and your own board will ask what it was told, what it was allowed, and what it did. Ancilla keeps that tamper-evident record at the checkpoint—ready before anyone asks.
A customer disputes what the agent did.
"Your agent sent the wrong amount." Your logs say one thing, the agent's log says another, and the customer has no reason to trust either.
A packet sets the agent's claim beside the signed decision, what the gateway forwarded and what the target system recorded. Where they disagree, the verifier flags it.
Your insurer or auditor asks how you control agents.
Companies seeking insurance on their AI, or certification such as AIUC-1, are asked how they account for what their agents do. Policies describe intent; they don't show what happened on a given day.
A packet is a record from the day itself, with its limits stated on it.
Something goes wrong and you need to know why.
When an agent's own account is the only one, an investigation starts from a claim. Researchers have shown that common agent tools let an agent delete its own traces when asked.
Ancilla records from the gateway, outside the agent, and checks the record against checkpoints the agent can't reach.
On agents deleting their own traces: Qin et al., “LLM Agents Can Easily Tamper With Their Own Traces”, arXiv 2609.30266, September 2026.