The question.
Long-running agent tasks fail when a chat transcript is treated as a workflow engine. Work needs explicit state, recoverable execution, and a clear boundary around actions.
The design combines a supervisor with narrowly scoped workers. A task should survive a process restart and remain inspectable without exposing private model reasoning.
The proposed approach.
- Represent work as a state graph with persisted checkpoints
- Use bounded tool permissions and structured task handoffs
- Make external writes idempotent and pause consequential actions for review
- Replay failures against a versioned task suite
Target architecture.
Task briefPlannerSpecialist agentsTool gatewayReview checkpointTask ledger
The difficult parts.
- Recovering after a partial external action
- Separating useful parallelism from duplicated effort
- Keeping task memory relevant and scoped
How to evaluate it.
- End-to-end task completion against a single-agent baseline
- Recovery under injected tool failures and timeouts
- Unauthorized-action rate, latency, and cost per completed task
Planned deliverables.
- A resumable reference workflow
- An inspectable task and tool-execution ledger
- A failure-injection evaluation suite
Design principles.
Give an agent a narrow responsibility before giving it more tools.
Observe decisions and evidence, rather than relying on hidden reasoning.