Reconciling what a delegated A2A task cost: runnable example, localhost, no keys

[fadnisnikhil]

When an orchestrator delegates over A2A, the task stream says whether the work finished. It says nothing about what the work cost, whether the provider's bill matches the delegation, or what is still unpaid. I built a small example that answers those for one job.

An orchestrator runs one job through two A2A v1.0 agents over HTTP on localhost, using @a2a-js/sdk:

• Code-fix agent (paid through an agreed obligation). Its agent card advertises an ATCN agent id. It runs an adapter on its own task stream: TASK_STATE_WORKING becomes "started", its test, lint and patch artifacts become signed evidence, and TASK_STATE_COMPLETED becomes "completion proposed". A policy evaluates the evidence and clears USD 90 to the agent plus a USD 10 fee (simulated).
• Search agent (knows nothing about ATCN, bills on its own). The orchestrator records the delegation using the A2A task id as the provider's job reference, then imports the agent's bill. The charge for this task matches. A charge from an earlier job matches nothing and is flagged as unmatched_charge instead of being added to the total.

The job ends in a signed closure that anyone can verify offline:

git clone https://github.com/fadnisnikhil/atcn.git
cd atcn && npm ci
npm run demo:a2a

roll-up
  net cost USD 112.00 (charged USD 102.00, fees USD 10.00)
  reported paid USD 90.00 (sandbox, simulated), unresolved USD 22.00
open exceptions: 1
  unmatched_charge: no task or delegation has a matching stable reference
offline verification
  VALID   task closure
  VALID   closure package of obl_...

Change any amount in task-closure.json and the verifier reports INVALID.

Example: https://github.com/fadnisnikhil/atcn/tree/main/examples/a2a-delegation

Limits: settlement is simulated (no money moves), all agents share one in-memory network in one process, and the policy checks the evidence the agent submits rather than running the code.

Two questions for people delegating paid work over A2A:

1. What do you use today as the stable reference between a delegated task and the provider's bill: the A2A task id, a context id, or something in metadata?
2. Would an A2A extension that declares "this agent bills against this reference" be useful, or does that belong outside the protocol?

If you have a project that delegates paid work, I'm happy to help wire one of your jobs into a reconciled record. Reply here or open an issue on the repo.

[kaspermuutra-dotcom]

On Q1: in what we’ve seen, the task ID is the only reference stable enough to bill against, because a contextId can span several tasks, so a provider billing per context can’t be matched one-to-one. A small extension where the agent declares “I bill against taskId” would make unmatched_charge detectable from the card alone, before any money moves. It would also be testable from outside: call the skill, then check the bill references the right task.

Kasper