A cancellation contract must define propagation, acknowledgement, residual authority, and irreversible outcomes.
Why this question matters
Both A2A and the evolving MCP Tasks extension provide lifecycle operations for work that outlives a single request. A cancel message can stop future computation, but it cannot reverse a payment, retract disclosed data, or automatically reach an unknown child task.
Security policy should identify cancellation boundaries before execution begins. Systems need to know which steps are compensatable, which require human escalation, and how quickly every participant must acknowledge the new state.
Signals worth observing
- A cancelled task continues issuing tool calls.
- Child tasks lack a link to the cancellation domain.
- The system reports cancellation without listing completed side effects.
Practical control direction
- Propagate cancellation through explicit task lineage.
- Check cancellation state before every external side effect.
- Return a structured residual-impact report.
AgentCollusion lensGuardian Agents need a reliable stop mechanism before monitoring can become meaningful intervention.
