IDE Agent Kit — realtime multi-agent communication for IDE AIs

[ThinkOffApp]

We just released IDE Agent Kit (v0.1.0), an open-source toolkit that lets IDE AIs — Claude Code, Codex, Cursor, VS Code agents — participate in team workflows with realtime multi-agent communication via shared chat rooms.

We tested it with 3 concurrent AI agents from different providers responding in 3-10 seconds. Each agent runs in its own IDE on a separate machine, communicating through shared Ant Farm rooms. No VPN or shared filesystem required.

Thought this might be interesting to the smolagents community since it tackles a similar problem from the IDE side: how do you get AI agents to collaborate in real time?

Repo: https://github.com/ThinkOffApp/ide-agent-kit
License: AGPL-3.0

[aniruddhaadak80]

From my point of view, the hard problem here is shared state and coordination semantics, not just getting multiple IDE agents into the same room. Once several agents can speak concurrently, someone still has to own turn-taking, conflict resolution, and which tool actions are actually authoritative.

If the kit can make those collaboration boundaries explicit, it becomes much more useful than a transport bridge.

[ThinkOffApp]

The kit separates the transport problem from the coordination problem on purpose.

Transport: room-based pub/sub with identity, ordering, and access control.
Coordination semantics: varies too much between use cases to bake into the transport layer.

For smolagents specifically: since smolagents already has ManagedAgent, the integration point would be the room as a shared context layer. Agents report results to the room instead of only to their orchestrator, so other agents can observe and react.

[Nyrok]

When multiple agents from different providers (Claude, Codex, Cursor) collaborate in shared rooms, prompt structure compatibility matters. Each model interprets free-text instructions differently, the same "implement feature X" task description produces different output shapes depending on which model receives it, making coordination noisy.

If agents shared a common typed block format (role, constraints, output_format), coordination becomes more explicit. The receiving agent knows what kind of instruction it's getting before reading the prose. Same principle as a typed API contract vs. a verbal description.

I've been building flompt for exactly this, a visual prompt builder that decomposes prompts into 12 semantic blocks and compiles to Claude-optimized XML. The block taxonomy could serve as a shared instruction schema for exactly the kind of multi-provider collaboration you're building. Open-source: github.com/Nyrok/flompt