What should count as one agent operation?

[roy-tong]

AgentMeasure needs a stable definition of an operation before its data can be comparable across providers and runtimes.

A concrete case:

1. An agent decides to use capability A.
2. 2. The first tool call times out.
3. 3. The runtime retries capability A.
4. 4. That attempt fails.
5. 5. The runtime falls back to capability B.
6. 6. Capability B succeeds and completes the user task.
      Possible measurement models:

• Model A: 1 operation, 3 attempts, 2 providers
• • Model B: 3 operations
• • Model C: 2 operations, one per provider
• • Model D: operation is defined at the user-task level, so all of this is one operation
    My current leaning is:

Operation = one agent-level intent to obtain a capability outcome.
Retries are attempts under the same operation; an explicit fallback to a semantically different capability may start a new operation.

But this breaks in several edge cases: parallel tool calls, speculative execution, handoffs, subagents, HITL pauses, and composite tools.

I'm looking for counterexamples from real runtimes.

If you maintain an agent framework, MCP server, observability system, or tool provider: what is the smallest boundary you would trust for cross-provider usage measurement?

A concrete trace or failure case is more useful than a general opinion.