Multi-Agent Orchestration Across Copilots

The Plasma, Engineering, and Operations copilots collaborate through L4 with clear ownership, shared grounding, and a single human authority.

Three specialists, one workflow

Real operator questions cross copilot boundaries. "Can we run this high-confinement scenario tonight?" needs the Plasma Copilot's stability assessment, the Engineering Copilot's magnet-margin check, and the Operations Copilot's runbook feasibility. Kronos orchestrates the three specialists rather than building one monolith, so each stays deep in its domain and grounded in its own evidence.

Orchestration through L4

Cross-copilot workflows are coordinated through L4, which routes a compound request to the relevant copilots, passes shared context (the same twin snapshot and session state), and assembles their contributions. Each copilot answers within its authority; none can invoke a side-effect on behalf of another. A proposal still passes the single authorization gate and single human approval.

orchestrate(compound_request):
  subtasks = decompose(compound_request)   # by domain
  results = {}
  for st in subtasks:
     copilot = route(st)                    # plasma/eng/ops
     results[st] = copilot.run(st, shared_context)
  merged = integrate(results)               # flag conflicts
  return merged   # one answer, one proposal path, one approver

Conflict is surfaced, not hidden

• Plasma says a scenario is stable; Engineering says magnet margin is thin
• Orchestrator surfaces the conflict to the operator rather than averaging it
• Each copilot's confidence and evidence travel with its contribution
• The operator resolves the trade with full context from all three

When copilots disagree, the orchestrator presents the disagreement explicitly — the whole point of separate specialists is that their independent, grounded views can be compared. Silently reconciling a conflict would destroy that value and could hide a real risk. This mirrors the founder-set principle that contradictions are surfaced for a human to decide, never quietly resolved.

Shared grounding keeps the specialists consistent: they read from the same L2 fabric and the same twin snapshot, so they are reasoning about one machine state, not three. The composed answer carries every copilot's citations and uncertainties into human oversight, and the full multi-agent trace is retained for audit.