[FEATURE] Durable Agent

[hertaehyun]

Problem Statement

Durable agents are AI agents designed for production, offering resilience, state management, and cost-efficiency by persisting progress across failures (crashes, restarts) and long waits, unlike ephemeral agents.

Temporal (https://docs.temporal.io/) and DBOS (https://docs.dbos.dev/why-dbos) provide durable capability. Also, Microsoft Agent Framework provides kind of durable capability(https://learn.microsoft.com/en-us/agent-framework/user-guide/agents/agent-types/durable-agent/features?pivots=programming-language-python)

The Strands Agents provides significant benefits when using a durable agent, including crash recovery, automatic retries, idle agent suspension, simplified development, and enhanced observability.

Proposed Solution

AWS Lambda durable functions simplify building resilient, long-running, multi-step applications and AI workflows directly within the existing Lambda programming model. They allow functions to automatically checkpoint progress, suspend execution for up to one year without incurring compute charges, and recover from failures, all without requiring you to manage additional infrastructure.

By integrating with AWS Lambda, I hope Strands can provide durable capability.

Use Case

Use cases and industries requiring complex multi-agent collaboration patterns—such as Graph, Workflow, and Swarm—will benefit

Alternatives Solutions

No response

Additional Context

No response

[krokoko]

Thanks @hertaehyun , I was going to open the same issue :)

To add to this:

It seems different SDKs provide native integration with durable execution solutions:

• Pydantic AI supports temporal, DBOS, Prefect
• OpenAI SDK supports temporal
• Vercel has the Workflows SDK

It would be great to have some native integration within Strands.

[zastrowm]

#1138 is related as I presume snapshots/checkpoints will play into the implementation.

This issue is a good one for tracking integration with different frameworks.

[jacques-]

We've run into ratelimits with agentcore runtime when trying to run into small numbers of parallel agent flows (~50) and through multiple tries from AWS side haven't been able to work through it to increase limits. The SE suggestion was to move off agentcore runtime to lambda (which we've now completed) and suggested we look at durable functions - so massive +1 on this 🙏 🙇

For our use-case, (beyond the reliability improvements) we'd like to have a way to manage the 15 min lambda runtime if the agent is running for short steps but triggering some fairly long running (~10 mins per call) tools.

[kosti-v]

@jacques- thanks for surfacing this. I am the PM for AgentCore Runtime, and I apologize for the inconvenience. We are working on making increasing limits on AgentCore Runtime easier (without manual intervention from us) as we speak. In the interim, mind emailing me the limit needed, your account id and region at kvasilak@amazon.com? We will expedite your limit increase.