Home โ€บ Blog โ€บ Durable execution
Inside the Agent Layer ยท Part 3 ยท Durable execution

Agents That Survive: Durable Execution for Long-Running Work

A chatbot answer takes seconds. An agent task can take hours: it calls a model dozens of times, runs tools, waits for a human to approve a step, and waits again for a CI build. Over that time servers crash, deployments restart processes, model APIs time out and rate limits bite. If the agent's progress lives only in one process's memory, all of it is lost. Durable execution is the infrastructure that makes a long agent run behave as if nothing ever failed: it records each step, and after a crash it replays the record and carries on from where it stopped.

FL
FrontierAGI Team
โš ๏ธ How to read this (as of 28 Oct 2026) Temporal's Series D is from its own announcement; the later Series E and customer figures are reported by secondary outlets (โš ๏ธ). Framework behaviour is from official documentation. The crash simulator and the failure calculator are our illustrations.

1. Case file

โณ
Durable execution for agents, October 2026Part 3 of the agent layer
Temporal Series D (Feb 2026)$300M at a $5B valuation (company announcement)
Temporal Series E (reported)$550M at $12.55B โš ๏ธ
OpenAI Agents SDK + TemporalIntegration built with OpenAI; public preview 2025
LangGraphCheckpoint every step; interrupt() for human approval
Other enginesRestate, Inngest, DBOS, AWS Step Functions and more

2. The problem

A naive agent is a loop in one process: ask the model, run the tool it chose, append the result, repeat. That works in a demo. In production, the process can die at step 37 of 60. Restarting from step 1 wastes the model calls already paid for, and worse, it can repeat side effects: a second email sent, a second payment, a second pull request. The question durable execution answers is: how do we restart a half-finished agent so that work already done is neither lost nor repeated?

3. How durable execution works

1. Split code into two kindsThe workflow is the deterministic plan (the agent loop). Activities are anything that touches the outside world: a model call, a tool call, an API request.
2. Record every resultEach time an activity finishes, the engine writes its result to a durable event history (a database) before the workflow moves on.
3. Replay after a crashA new worker re-runs the workflow code from the start, but instead of calling activities again it reads their recorded results, until it reaches the first step with no record.
4. Retry what failedActivities that failed or timed out are retried with back-off by policy, so a flaky model API or a 429 rate limit becomes a delay rather than a dead run.

The trick is that model outputs are not deterministic, so they must be recorded rather than recomputed. Once the model's answer for step 12 is in the history, replay sees exactly the same answer, and the agent takes exactly the same path.

4. Crash and replay

Kill the worker at step 5

A seven-step coding-agent task. Pick a strategy and press run; the worker crashes after step 4. Our illustration.

5. Why it matters more as tasks grow

Chance a run hits at least one failure

Our model: 1 โˆ’ (1 โˆ’ p)n, with n external calls and p the chance that any one call fails transiently (timeout, rate limit, crash).

Task length is rising fast (Evaluation Layer Part 5 tracks the time horizons). An agent that works for a working day makes hundreds or thousands of calls; at that scale a failure somewhere is close to certain, so recovery has to be built in, not bolted on.

6. The tools

ToolApproachAgent angle
TemporalOpen source (MIT) workflow engine with event-history replay; Temporal Cloud is the paid serviceIntegration with the OpenAI Agents SDK built with OpenAI; OpenAI says it uses Temporal for durable orchestration; reported to be Temporal's biggest customer โš ๏ธ
LangGraphGraph of nodes; a checkpointer saves state per thread after each stepDurability modes exit / async / sync; interrupt() pauses for human input; Postgres checkpointer for production
Restate, Inngest, DBOSLighter-weight durable functions: journal steps in a log or in PostgresPitched at AI teams who want durability without running a separate cluster
Cloud workflow servicesAWS Step Functions, Azure Durable Functions, Google Cloud WorkflowsManaged state machines; often used to orchestrate agent steps inside one cloud
Agent platformsHosted agent runtimes from the labs and cloudsPersist session state and resume; details are mostly not public

7. Waiting for humans

Durable execution is also what makes human-in-the-loop practical. When an agent needs approval to merge code, spend money or send a message, it should not hold a process and a sandbox open for the hours a person may take to answer. Instead it saves its state and stops. When the person approves, the engine loads the state and resumes. In LangGraph the node calls interrupt() and is resumed with a Command; in Temporal the workflow waits for a signal, which can take days at almost no cost. One caveat from LangGraph's docs: on resume the interrupted node re-runs from its start, so code before the interrupt must be safe to repeat.

8. Rules for agent builders

Make tools idempotentPass an idempotency key with every side-effecting call (payment, email, ticket) so a retried call is recognised and not repeated.
Record model outputsNever re-sample the model during replay; store its answer as the step result.
Keep the workflow deterministicNo random numbers, wall-clock reads or direct network calls in the workflow code; put them in activities.
Version carefullyA run that started on Monday may replay on Thursday's code. Engines offer workflow versioning so old runs follow old logic.
Watch history sizeLong agents with big contexts produce large histories; summarise or store large payloads outside, and link (see memory, Part 6).

9. The series

Inside the Agent Layer: eight parts

VerdictDurable execution turns a long agent run from a fragile process into a recorded, resumable workflow: every model and tool result is written down, crashes are replayed, failures are retried and humans can take days to answer. As tasks stretch to hours, this is not optional. But an agent that survives still needs to reach the outside world, through tools. Part 4 covers tool gateways and MCP at scale.

10. Reading the numbers

Primary
Temporal Series D, $300M at $5B (Temporal blog, Feb 2026). LangGraph checkpointers, durability modes and interrupts (LangChain docs). Temporalโ€“OpenAI Agents SDK integration (Temporal announcement).
Reported โš ๏ธ
Temporal Series E, $550M at $12.55B, and OpenAI as its largest customer (Forkast). GA date of the OpenAI integration (Unite.ai).
Our illustrations
Crash simulator, failure-odds calculator.