Agent teams in deep-agents

[Jonathan Reynold Gomes (devgomesai)]

Hey devs! 👋
I was going through the differences between Sub-Agents and Agent Teams — and honestly, Agent Teams deserve to be a first-class feature in deep-agents.
Sub-agents are great, but Agent Teams unlock:
Real parallel collaboration
Role specialization (planner, executor, critic)
Shared context across agents
It's a natural fit with LangGraph's graph-based design and would make complex, long-horizon tasks way more manageable.
Would love to see this on the roadmap — what do you think? 🚀

References: https://code.claude.com/docs/en/agent-teams

[krampus (krampusx64)]

Totally agree that Agent Teams should be first-class here. The planner/executor/critic pattern in particular is where things get interesting — but the hard part isn't the architecture, it's the shared context management. When three agents are working in parallel and each modifies the shared state, how do you avoid one agent's stale read invalidating another's output?
LangGraph's graph design handles sequential dependencies well, but parallel writes to shared context is a different problem. Has anyone explored using a message-passing model between team members instead of a shared mutable state — more actor-model style? Curious if that would actually make long-horizon tasks more reliable or just add coordination overhead.

[Manan Jariwala (itsmanudon)]

It seems to be difficult to do this. Race conditions are very common in such architecture. Multiple agents may run but they need to perform different tasks or else one may override the work of other.

It's the same scenario as git tree, one developer can override the complete commit history.

For such architecture to work, all the agents should work on shared resources together.