Durable Agents Explained: Surviving Failure
Temporal raised $550M at $12.5B for durable execution. Why long-running AI agents fail, what retries miss, and a checklist to ship agents that finish.
On this page
Demos forgive failure, production bills for it. On September 14, 2026 Temporal raised $550M at a $12.55B valuation on one thesis: agents die mid-run, and retries alone do not save them. Here is what breaks in long-running agents and how to build agents that finish without adopting a whole platform on day one.
Takeaways
Agent runs fail multiplicatively: every tool call, model call, and API hop adds a chance to crash, time out, or hang. Durable execution means resume, not retry: persisted step state plus idempotent tools beats naive re-runs. Temporal's numbers are the signal: 1.9T actions in August, 4,300 customers including OpenAI: long runs are now normal. Solo pattern: checkpoint to Postgres, idempotent tools, bounded retries, durable queue for anything over 60 seconds.
Why do agents fail in production?
Three failure shapes dominate, all boring, all expensive:
- Crash mid-plan. The process dies at step 7 of 12: API timeout, OOM, deploy restart. Naive retry re-runs steps 1 to 6, double-charging, double-posting, double-emailing.
- Hang forever. A tool never returns, the model waits, the user waits. No timeout, no lease, no escalation: just a stuck run holding money and attention.
- Partial writes. Step 4 wrote to the DB, step 5 failed, retry writes step 4 again. Without idempotency keys, every recovery corrupts state a little more.
demo run: 12 steps x 99% reliable = 89% success, nobody notices
prod run: 200 steps x 99% reliable = 13% success, pager fires
fix: checkpoint each step + idempotent tools + resume from last goodThe Lightspeed line worth stealing
"The demo is easy, production is hard, because the systems around the models can't handle real-world execution." That is the whole pitch: models got good enough that the plumbing is now the product.
What does durable actually mean?
Persist step state, not just logs
After every tool call, save inputs, outputs, and next intent to durable storage: Postgres row, not memory. A new process must be able to pick up exactly where the dead one stopped.
Make every tool idempotent
Same key, same effect twice. Pass an idempotency key into writes, emails, charges, and file ops: then retries are safe by construction instead of scary by default.
Bound retries with backoff and leases
Three retries with exponential backoff, per-step timeouts, and a lease so only one worker owns a step. Inngest, Trigger.dev, and Restate all sell variants of this: the pattern ports anywhere.
Keep a human-readable trace
Every resume needs an audit: what ran, what failed, what was skipped. A simple append-only run log turns "the agent did something" into something you can debug at 2am.
When should you reach for Temporal?
Use the revenue test from this week's news: Temporal crossed $250M run-rate with net retention over 200% because customers expand once agents touch payments and support. Reach for durable infra when runs exceed minutes, money moves, or failures page humans. Below that, Postgres plus a queue plus idempotency keys ships fine: the same stack behind most indie agent tooling.
What breaks if you skip this?
My agent runs are short: can I skip durability?
If every run finishes in under 60 seconds with no side effects, yes. The moment you add webhooks, long tool chains, or scheduled follow-ups, add checkpoints first, that is where the 2am pages come from.
Is Temporal the default choice now?
It is the best-funded choice, not the only one. Inngest for event-driven flows, Trigger.dev for TypeScript long tasks, Restate for lightweight durable code, or plain Postgres checkpoints for solo projects: all implement the same resume-not-retry idea at different scales.
As of September 16, 2026: the $550M round prices one lesson: agent capability stopped being the bottleneck, completion did. Checkpoint each step, key every write, bound every retry, and your agents finish like infrastructure instead of expiring like demos. Next, read who pays for the AGI race for where that infra money flows.