What is a multi-agent control plane?
A control plane is the operator layer that assigns work to specialized agents, shares tools through a fabric such as MCP, and records every hand-off. Without it, “multi-agent” is several chat sessions racing to the same repo. With it, roles are explicit, risky steps wait for a human, and you can reconstruct who did what.
Proof from Conductor and Nexora
Conductor runs an operator UI on Cloud Run: Hermes plans and verifies, coding agents execute against repos, MCP is the shared tool fabric, kanban and delegation briefs stay durable, and auto-merge to main stays off. Nexora is the single-operator cousin—one HITL IDE where independent DAG steps run in parallel after a human approves the plan.
How we design the fleet
- Role charter. Write who plans, who executes, who verifies, and what a human must still approve. If two agents can both merge, the design has already failed.
- Tool fabric. Expose repos and ops tools through a shared protocol (MCP or equivalent) so agents do not each get a private, unbounded credential.
- Intake and briefs. Structure work as briefs the orchestrator can assign—CLI, cron, or ticket—not as a freeform chat that evaporates.
- Audit and gates. Append-only logs, an approvals queue, and a hard rule that production branches do not merge unattended.
What you receive
- Role map and runbooks for orchestrator, executors, and humans
- Control-plane UI or API: queue, briefs, approvals, audit trail
- Tool-fabric wiring (MCP or equivalent) with least-privilege credentials
- Policy: which tasks may run in parallel and which require a gate
- Handoff so your engineering ops team can operate the plane
Related work and reading
Frequently asked questions
Why not just run several coding agents in one chat?
They collide: duplicate PRs, silent overwrites, no audit. A control plane gives each agent a role, a brief, and a gate. Conductor exists because ad-hoc multi-agent coding produced unattended merges.
Does VisionsCraft auto-merge agent work to main?
No. Conductor keeps auto-merge to main off by design. Humans approve high-risk or production-impacting work. That is a product decision, not a missing feature.
What is MCP used for?
Model Context Protocol is the shared tool fabric: the orchestrator and coding agents reach repos and operational tools through one permissioned layer instead of scattering API keys.
Is this only for software engineering?
Engineering is where we have the deepest proof (Conductor, Nexora). The same role-split pattern applies to GTM fleets such as Schmoozzer, where outreach channels are specialized but pipeline state stays one system.