# Durable execution

Execution that persists state and can resume safely across process, model, context, worker, or environment interruptions.

technical-term · Harness & runtime · maturing · Reviewed 2026-08-09

## Definition

Durable execution externalizes an agent’s state transitions, checkpoints, side-effect status, budgets, and evidence so work can resume without replaying unsafe actions or depending on one model context. It applies distributed-systems principles such as idempotency, leases, retries, and compensation to agent runs.

Workflow durability is established; model-driven, stateful agent runtimes are still evolving.

## Why it matters

Long-running agency is state-machine engineering as much as model prompting.

Resumption must distinguish work that was proposed, started, committed, verified, or compensated.

## System anatomy

- **Checkpoint:** A durable snapshot of task and control state.
- **Idempotency:** Protection against duplicate effects during retry or recovery.
- **Lease:** Time-bounded ownership of work by one executor.
- **Compensation:** A defined way to reverse or remediate partial effects.



## Important distinctions

- **Long context window:** More tokens do not provide transactional durability, concurrency control, or safe replay.
- **Conversation persistence:** Saving messages does not capture exact side-effect and verifier state.



## Implementation signals

- Persist state outside model context
- Make external effects idempotent or compensatable
- Checkpoint verifier and approval dispositions

## Failure modes

- Recovery repeats a payment, deployment, or message
- Stale workers continue after lease expiry



## Related knowledge

- [Working state](https://aisdlc.ai/agentic-engineering/working-state) — The current, inspectable record of an agent’s task position, observations, decisions, pending actions, and budgets.
- [Agent loop](https://aisdlc.ai/agentic-engineering/agent-loop) — The finite runtime cycle in which an agent interprets state, selects an action, observes the result, updates state, and decides what happens next.
- [Harness engineering](https://aisdlc.ai/agentic-engineering/harness-engineering) — Engineering the agent loop, task decomposition, tools, permissions, session state, checks, retries, feedback, checkpoints, and stop conditions that surround a model.
- [Agent execution substrate](https://aisdlc.ai/agentic-engineering/execution-substrate) — The isolated, stateful environment in which agents observe and act, including compute, filesystem, browser, network, credentials, resource limits, and session lifecycle.

## Sources and further study

- [Anthropic — Scaling Managed Agents: Decoupling the brain from the hands](https://www.anthropic.com/engineering/managed-agents) — An architecture separating durable sessions and environment state from changing models and harnesses. **Use:** First-party case study. **Limitation:** This first-party account documents one organization, product, or implementation context and should not be generalized without local evidence.
- [Anthropic — Effective harnesses for long-running agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) — Patterns for incremental progress and durable handoffs across context windows and extended execution. **Use:** First-party case study. **Limitation:** This first-party account documents one organization, product, or implementation context and should not be generalized without local evidence.
- [OpenAI — An open-source spec for Codex orchestration: Symphony](https://openai.com/index/open-source-codex-orchestration-symphony/) — An orchestration specification that connects project work, isolated agent execution, and reviewable delivery state. **Use:** First-party case study. **Limitation:** This first-party account documents one organization, product, or implementation context and should not be generalized without local evidence.

---

This library synthesizes cited research, standards, official documentation, and clearly attributed practitioner perspectives. Maturity describes the state of a concept—not vendor endorsement, production readiness, or permission to deploy. Benchmarks and demonstrations do not replace use-case evaluation, governed controls, independent verification, or named human release authority.