FIELD NOTE / X
Agent tasks finished is a weak score.
The short film, the complete written thought, and the evidence behind it.
The X conversation link will follow its public release.
Agent tasks finished is a weak score.
Video caption
Agent tasks finished is a weak score. Time to known state belongs beside throughput. Grant autonomy to classes with a measured safe terminal path. #EricFieldNotes
Full written post / accessibility read
A team celebrates a rising count of autonomous service moves. The operational queue of unknown placements and manual reconciliations grows at the same time. The first metric hides the second.
Track unknown-effect backlog, time to reconcile, customer-visible reversals and operator hours. Slice by action class. A low attempt cost means little if recovery dominates.
Use a disposable control-plane fixture to inject those four cases. Require the workflow to halt, reconcile or escalate correctly without a second unapproved effect.
Expand agent scope only when both accepted work and recovery time meet the team's threshold. Do that because activity is cheap to produce while an unresolved effect is expensive to own.
#EricFieldNotes
Four-beat scene transcript
1. Agent tasks finished is a weak score.
A team celebrates a rising count of autonomous service moves. The operational queue of unknown placements and manual reconciliations grows at the same time. The first metric hides the second.
Visual: Unresolved effects accumulate after the green turns.
2. Measure the path back.
Track unknown-effect backlog, time to reconcile, customer-visible reversals and operator hours. Slice by action class. A low attempt cost means little if recovery dominates.
Visual: Time to known state belongs beside throughput.
3. Run interruption drills.
Use a disposable control-plane fixture to inject those four cases. Require the workflow to halt, reconcile or escalate correctly without a second unapproved effect.
Visual: Lost reply, crash, partial fanout and stale compensation.
4. Scale on recovery performance.
Expand agent scope only when both accepted work and recovery time meet the team's threshold. Do that because activity is cheap to produce while an unresolved effect is expensive to own.
Visual: Grant autonomy to classes with a measured safe terminal path.
Research and claim limits
- AWS Step Functions redrive (S214)
- Azure Compensating Transaction pattern (S215)
- OpenTelemetry messaging spans (S217)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.