Lambda durable functions quotas: what changes for agents

On November 5, 2026, AWS raised the default quotas for Lambda durable functions: running executions went from 1 million to 5 million per account in most Regions, and the rate limits on starting and checkpointing executions went up tenfold. If you run, or plan to run, multi-step AI workflows on Lambda, this removes a ceiling you would have hit at scale. If you run a few thousand executions a day, you can ignore it.

What was announced, and what is different?

AWS's announcement lists the new defaults. They apply in all Regions except US East (N. Virginia), US West (Oregon) and Europe (Ireland), which get twice these values:

• Running durable executions per account: 1 million to 5 million (the quotas page says per Region).
• Function invocation rate to start durable executions: 300 to 3,000 requests per second.
• CheckpointDurableExecution and GetDurableExecutionState: 1,000 to 10,000 requests per second.
• Callback success, failure and heartbeat calls: 300 to 3,000 requests per second.
• GetDurableExecution and StopDurableExecution: 30 to 300 requests per second.
• GetDurableExecutionHistory and ListDurableExecutionsByFunction: 15 to 150 requests per second.

The Lambda quotas page now shows the same numbers, with 10 million running executions in the three Regions with higher defaults. The quotas in the list above are soft limits you can raise through Service Quotas. Other limits are not adjustable, and the announcement does not list them as changed: 30,000 durable operations per execution, 100 MB of persisted data per execution, and 1 MB per operation payload.

Before1 million

After5 million

Source: AWS What's New, November 5, 2026, for Regions other than N. Virginia, Oregon and Ireland.

What are durable functions, in one minute?

Durable functions are regular Lambda functions that save progress as checkpoints. If the function is interrupted, or paused on purpose, Lambda replays your code from the start and skips the steps that already finished, using their stored results. An execution can last up to one year. Waits suspend the function without compute charges.

You write the workflow as ordinary code in a supported language (the SDK covers JavaScript, TypeScript, Python, Java, Go and C#). Each unit of work goes inside a step. AWS lists AI workflows that chain model calls, wait for a human, and resume as a target use.

Replay re-runs your code from the top and reuses saved results for finished steps.

Who should care about this change?

You should care if:

• You run agent or document pipelines with many executions waiting at once, for example on human approval or a slow external system. A wait can last a long time, so many executions can be "running" at once even when little compute is in use.
• You start executions in bursts. The old limit of 300 starts per second was a real ceiling for a batch job or a queue that drains all at once. It is now 3,000 (6,000 in the three Regions with higher defaults).
• You build dashboards or scripts that poll execution status. AWS says they can now check status, list executions, read history and stop executions at least ten times as often.

You can ignore it if your volume is low, or if you have not adopted durable functions. A quota is a ceiling, not a target.

What does it cost, and what does not change?

Quotas do not change the price. As of November 2026, the Lambda pricing page bills durable functions on four things: standard compute, durable operations, data written and data retained. In its US East (N. Virginia) example, durable operations cost $8.00 per million, data written $0.25 per GB, and retention $0.15 per GB-month. Retention after completion is configurable from 1 to 90 days, with a default of 14.

Each step, wait and callback creates operations, and a retried step counts again. AWS documents the metering per operation type. A quick calculation shows the scale. Take an illustrative workflow of 25 durable operations (not a benchmark). At $8.00 per million, that is about $0.0002 in operation charges per execution, before compute and data. For one million executions, that is roughly $200. Model tokens may cost more than the orchestration layer, so size those too.

A bigger quota also means a bigger blast radius. A runaway loop that starts executions can now do so ten times faster before the start-rate quota throttles it.

How do durable functions compare with Step Functions?

AWS's comparison guide is short. Choose durable functions when your logic lives mostly in Lambda and your team prefers code and its usual tests and tools. Choose Step Functions when you want a visual workflow, native integrations across many AWS services, or non-technical people to read the flow. AWS also describes a hybrid: durable functions for application logic, Step Functions for coordination across services.

Nothing in this quota change shifts that decision. It only means that scale is less likely to be the reason you pick one over the other.

How would a team try it safely this week?

1. Pick one low-risk workflow, such as a nightly document summary, not a payment flow.
2. Wrap each model call and each external write in its own step, so a replay never repeats a finished call. Keep step names stable: AWS warns that renaming a step while executions are in flight can stop them from resuming.
3. Invoke through a version or alias, not $LATEST, as AWS recommends.
4. Attach a dead-letter queue for asynchronous invocations. AWS's best-practices page covers steps 2 to 5: it warns about renaming steps, recommends versions or aliases, and notes Lambda does not retry failed durable executions on its own.
5. Add CloudWatch alarms on error rate and duration, and an EventBridge rule for FAILED, STOPPED and TIMED_OUT.
6. Load test above your expected peak (our rule of thumb), and watch throttling on the checkpoint API before you widen the rollout.

What would we check first?

Before moving anything to production, we start with four questions:

• Are all side effects idempotent? Replay is safe for code inside steps only if retries do not double-charge a card or send a message twice.
• How large are the payloads? Large model outputs count toward the 1 MB per-operation and 100 MB per-execution limits, and toward data written and retained.
• What is the real peak? Compare it with the start rate and running-execution quotas for your Region, using the Service Quotas console.
• Is the Region right? Check that durable functions are available where your data must stay, including for Canadian residency needs.

If you want a second pair of senior eyes on an agent architecture, our services page describes how we work.

This article is general information, not financial or architecture advice. Check current AWS documentation and pricing for your Region before you decide.