Distributed, durable execution for Swarm and Graph multi-agent patterns?

[sirianni]

I've read through the docs and various GH issues and PRs

• feat(a2a): configurable request handler #601
• [FEATURE] Support durable task store and event queue #532

It seems that the pluggable (durable) task store and event queue are only usable with A2A and not for programmatically defined flows (Swarm, Graph, Workflow). It's also not clear how "shared state" should be accessible to distributed workers.

Can someone share any insights as to the longer-term vision here?

[sirianni]

Wile this seems like a plausible response on the surface, looking at your GitHub profile you seem like a bot. What is your affiliation with the Strands project and AWS? I don't see any prior contributions from you in this repo and it looks like you just joined GitHub three months ago...

[yonib05]

Hi @sirianni, thanks for the thoughtful question, and apologies both for the noise earlier in this thread and for how long it took to get you a real answer. This one deserved better.

The picture has improved since you asked. Durable execution for the programmatic patterns landed: Graph and Swarm now accept a session_manager (file, S3, or your own implementation) that checkpoints the orchestration state and lets you resume after a crash or restart. Only the orchestrator holds the session manager; giving an inner agent its own raises a ValueError (nested session managers aren't supported yet). The orchestrator's session manager persists the orchestration state and each node's results, though a node that was mid-flight at the crash re-runs on resume rather than resuming partway. Docs are here: session management.

You're right that the pluggable task_store and queue_manager are A2A-server concepts and stay there. They cover the protocol-level task lifecycle, while the session manager covers Graph/Swarm execution state, so the two are complementary rather than one being a subset of the other.

On shared state: invocation_state is a plain dict passed into the invocation and propagated in-process to nodes and tools. It isn't persisted or synchronized, so with distributed workers anything shareable needs to live in an external store you manage (the dict is a fine place to pass handles to that store).

True distributed execution of a single Swarm or Graph across workers isn't built in today. The supported shape is one orchestrator process that can be rehydrated on any host via its session manager, with heavy agents scaled out separately behind A2A servers. Deeper durable-execution work is being tracked in the open issues, for example #3662 (snapshot-style persistence for Graph/Swarm) and #2243 (tool-level suspend/resume for external async callbacks). If one of those matches what you need, give it a 👍. We use those to figure out what to build next. And your Temporal/DBOS experience would make a great comment there too.