Why swarms are the hard case
One agent is a line of calls. A swarm is dozens of agents acting at once: scouts gathering data, workers processing in parallel, a coordinator merging results, sub-swarms spawning mid-run. The final output is a blend of a hundred partial contributions, and the usual record is the coordinator's summary.
Summaries fail swarms completely. When the output is wrong, no summary says which of the forty workers introduced the error. When the bill arrives, no summary says which branch burned the budget. When a regulator asks what happened, forty disconnected agent logs do not compose into an answer.
The evidence chain pattern
- Receipt per call, per agent. Every tool call by every agent mints a verifiable receipt at execution time: tool, caller identity, timestamp, inputs, outputs, SHA-256 commitment.
- Parent linkage across agents. When agent A delegates to agent B, B's receipts reference A's delegation receipt. The chain crosses agent boundaries, so the tree spans the whole swarm.
- Merge receipts. When the coordinator combines partial results, it mints a receipt for the merge that references every input receipt. The blend itself is evidenced.
- Parallel verification. The tree verifies branch by branch, independently. A broken commitment flags exactly the agent and call that was tampered with; a gap flags a missing sub-run.
What the chain gives you
- Attribution. Every fragment of the final output traces to the agent and call that produced it.
- Fault isolation. A bad result walks back to the exact worker and call, not to the swarm as a whole.
- Cost decomposition. Per-receipt costs roll up by agent, branch, and phase, so the expensive part of the swarm is visible.
- Completeness. Sequence gaps show dropped workers and lost sub-results, which in a swarm is the difference between a partial answer and a wrong one.
Smaller version: orchestration trees
The same chaining pattern audits any multi-agent run, even with just a handful of agents. Read the orchestration audit guide.
AER-1 and swarm-scale evidence
AER-1 is a verifiable execution receipt format, published as the IETF Internet-Draft draft-zambo-aer1. Each receipt carries the tool, the calling agent, the timestamp, canonical inputs and outputs, a SHA-256 commitment, and linkage fields for parent references. Independent implementations mint these receipts, and the public conformance vectors let anyone verify a swarm's chain without trusting its operator.
The format is deliberately small: a receipt is cheap to mint and cheap to check, which is what makes per-call evidence affordable at swarm scale. Evidence that costs more than the work will never be adopted; evidence cheaper than the work becomes the default.
See the primitive in action
One receipt is the whole pattern in miniature. Mint a real AER-1 receipt for a real tool call, free, no signup, then picture every agent in your swarm minting one per call.