Execution plane / Tier 1 / Phase 1
Orchestration & Durable Execution
Who acts, when and in what order?
Plan and decompose the task, route work to the right agent or tool, checkpoint durably so long tasks survive restarts, and stop cleanly when budget or policy says so.
- Planning and decomposition; orchestrator-worker and handoff patterns
- Durable checkpoints, retries and idempotency for side-effecting steps
- Termination policy: step limits, budget ceilings, loop detection
- State machines that make agent behaviour inspectable
- Multi-agent topology decisions, including when not to use one
- Resumption after human approval without replaying side effects
only cleared requests enter the loop
grounds the task
requests context
calls the model
executes actions
writes back what is worth keeping
approval checkpoints and autonomy tiers
trace, thread and task correlation
delegation to peer agents
AI Engineering
- Principal agent architect — accountable owner
- Agent engineer — patterns, planners, handoffs
- Distributed systems engineer — durability, retries, idempotency
- SRE — termination policy, circuit breakers, on-call
Architectural recommendation, not a research finding.
Pilot needs a single orchestrator with durable checkpoints and a hard step and budget ceiling. Multi-agent topologies only after a single agent has been measured and found insufficient.
Signals to watch
- Task success rate and steps per successful task
- Token cost per successful task, not per call
- Circuit-breaker and termination trips
- Checkpoint recovery success rate
- Share of tasks resumed correctly after approval
- Infinite loops and runaway token spend without termination policy
- Non-idempotent retries producing duplicate real-world side effects
- Multi-agent architectures used by default: roughly 15× the tokens of a single chat turn, with token use explaining about 80% of performance variance
- Token duplication across agents (one framework measured at 72%) and plan-and-execute patterns cutting spend by around 83%
- Long tasks losing all progress on a restart
Tool and standards snapshot · last reviewed approximately mid-2026
Open source
- LangGraph
- Temporal
- Restate
- Microsoft Agent Framework (Semantic Kernel plus AutoGen)
- Google ADK
- OpenAI Agents SDK
- CrewAI
- Ray
- Airflow or Dagster for batch paths
Managed
- Managed durable-execution services
- Cloud agent orchestration runtimes
Standards
- A2A task lifecycle for delegated work
- Idempotency keys on all side-effecting calls
- OpenTelemetry span conventions for agent runs
Parallel exploration is worth roughly 15× the tokens only when the work genuinely fans out. Gate multi-agent expansion on measured return, not on architectural fashion.