Many agent systems look autonomous during a successful demo. The real test begins when a worker exits early, a tool times out, a handoff goes missing, or two parts of the system disagree about the current state.
The hidden failure mode
An agent can produce a convincing message without completing the work. It can edit some files, run some tests, and disappear before recording what happened.
If the orchestration layer treats the message as truth, it will report progress that the system cannot prove.
That is not primarily a model problem. It is a control-plane problem.
What a durable supervisor owns
A reliable supervisor should own the lifecycle state:
Queued → Spawned → Verifying → Complete / Failed / Blocked
The worker should provide a structured handoff. The parent supervisor should independently verify it and own the authoritative ledger transition.
The handoff should include:
- exact files changed;
- commands run and results;
- blockers;
- schema or configuration notes;
- artifact location;
- explicit exit state.
Failure handling is part of the design
The supervisor needs deterministic behavior for:
- worker disappearance;
- stale leases;
- missing handoff artifacts;
- failed ledger writes;
- failed next-task spawns;
- duplicate completion events;
- verification failures;
- wrong working directories;
- lost notifications.
Retries should be bounded. A system that retries forever is not autonomous; it is hiding failure.
Why the ledger must be parent-owned
Workers can write partial state, claim completion too early, or exit after changing the repository. Their output is evidence, not authority.
The parent control plane should:
- record the spawn;
- watch the worker status;
- validate the handoff artifact;
- run independent verification;
- append the lifecycle transition atomically;
- only then report completion or start the next slice.
This is the same principle used in commerce control: claims about an action must be connected to verifiable evidence.
The practical lesson
Agent orchestration is not just task decomposition. It is state management, evidence management, failure recovery, and clear authority.
The best system is not the one that never encounters failure. It is the one that detects failure quickly, preserves the truth about what happened, and recovers without asking a human to reconstruct the entire run.