Managed Agents: specialist kill / interrupt semantics not documented

Summary

Per the public Managed Agents docs (platform.claude.com/docs/en/managed-agents/overview) and third-party deep-dives, the coordinator can interrupt/steer specialist subagents via the "Steer or interrupt" lifecycle step. However, the kill semantics for individual specialists are not explicitly documented: what happens to a running specialist when the coordinator decides to abandon a path, what happens to its in-flight tool calls, whether its session-hour billing stops immediately or at next checkpoint, and whether its filesystem writes are rolled back or persist.

Motivation / Use Case

Customers evaluating Managed Agents for production workloads (multi-agent dispatch with cost discipline, regulated-data workloads) need to reason about failure handling. Without documented kill semantics, the operational model is hard to size — particularly for cost-conscious users where an orphaned specialist could keep billing $0.08/hr until session timeout.

This concretely surfaced during a 2026-05-22 vendor evaluation by a CNQ/Opticca team where Managed Agents was being compared to Claude Code's Agent tool + manual 4-cap discipline. The lack of documented kill semantics was one of three doc gaps that made the cost / control story hard to commit to.

Suggested Behavior

Add to the official docs:

1. Explicit description of how coordinator -> abandon specialist works (verb, lifecycle event name, propagation latency).
2. Billing behaviour: does the specialist's session-hour clock stop at the abandon event, or continue until its next checkpoint / timeout?
3. Filesystem behaviour: are in-progress writes by the abandoned specialist preserved in the shared container FS, or rolled back / quarantined?
4. Tool-call behaviour: in-flight MCP / tool calls — interrupted? Allowed to complete and discard the result?

A short matrix at the bottom of the lifecycle docs would resolve this for evaluators.