Skip to main content
When you run many agents on the same project at once, they need to coordinate without colliding — but for safety they run in isolated, egress-denied pods and can’t talk to each other directly. AgentBees provides a mediated coordination layer called the hive: every message flows pod → orchestrator → siblings, never pod-to-pod.
Everything a sibling agent produces is treated as untrusted data by the receiver. Agents are instructed never to follow instructions embedded in another agent’s messages — coordination carries information, not commands.

​
The hive tools

When coordination is enabled for your org, agents get a toolset (an MCP server): You can hint the agent to use these from the New Task form’s Agent coordination strip, which appends coordination guidance to your prompt.

​
Deterministic conflict detection

Beyond agent-to-agent chat, the orchestrator computes conflicts and diffs deterministically (no LLM): it fetches sibling task-<id> branches and runs git merge-tree to find conflicting paths, and git diff --numstat for pairwise diffs with an overlap list of files both branches touched. This powers the hive_conflicts tool and the optional Workspace Coordinator agent — a peer that never edits code; it only detects conflicts and tells each worker to fix its own side.
Pipelines vs coordination. A pipeline runs stages sequentially on one shared branch. Coordination is for parallel tasks that each have their own branch and need to avoid stepping on each other.

​
Cross-agent handoff

If a task pauses, you can resume it with a different agent CLI in place — on the same live pod and worktree, with no new workspace provisioned.
1

On a paused task, pick a resume agent

The resume banner offers Claude / Codex / Gemini.
2

AgentBees builds a handoff brief

A portable, agent-neutral summary — the task, the original request, where the previous agent left off, the uncommitted diff, and project notes.
3

The new agent takes over

The pod swaps the CLI in place and the chosen agent continues from the brief.
Handoff is great when one model gets stuck — hand the exact same in-progress workspace to a different agent and let it continue, without losing any state.