Scenarios
4
AgentOps ยท supervisor and workers
agentops-multiagent-supervisor
Adding workers multiplies coordination failure modes. Cycles, delegation depth, and partial worker failure need first-class outcomes.
Public · synthetic demoThe problem
A supervisor can deadlock or decompose forever even when every worker is simple.
Routing by role is only the happy path. A usable coordinator must stop cyclic waits, bound recursion, preserve partial results, and show which worker failed.
Input
Scripted task types exercise distinct coordination paths.
normal_pipeline delegation_cycle runaway_decomposition worker_error_handled Bus: FIFO State: namespaced blackboard Workers: deterministic mocks
The money shot
Two guardrail interventions stop the unsafe coordination paths.
Bus messages
12
Interventions
2
Worker errors preserved
1
The cycle stops before any worker runs. Runaway delegation returns 5 completed results, then stops when depth 6 exceeds cap 5. The flaky worker's error is aggregated beside 3 successful results.
How it's verified
Each scenario records status, result count, errors, and guardrail.
| Scenario | Status | Results | Guardrail |
|---|---|---|---|
| normal_pipeline | ok | 3 | none |
| delegation_cycle | guardrail_stop | 0 | cycle_detected |
| runaway_decomposition | guardrail_stop | 5 | max_delegation_depth |
| worker_error_handled | completed_with_errors | 3 | none |
Honest limitations
This demonstrates coordination plumbing, not collective intelligence.
- Supervisor and workers are scripted mocks with fixed decomposition.
- The bus is an in-process list with no persistence or concurrency.
- Budgets and cycle checks cannot judge factual quality.
- Real model workers add non-determinism, latency, and cost absent from this run.
Coordination failures remain visible instead of becoming silent hangs. jigonyoo.com · Back to hub