Orchestration at 10³–10⁴

A coding harness is a worker. A supervisor is a topology, a budget, an allowlist, and a human-on-the-loop router. Confusing the two is how a ten-agent demo becomes a Gartner cancellation. Mesh channel count is n(n−1)/2. Helix refuses mesh above 80 agents. Hub-and-spoke is the default to ~800. Above that the hub is the incident — hierarchy, with a supervisor per ~80 workers and a root per ~640.

Budget is two numbers: $0.75 per session (override in policy, never in the prompt) and an hourly fleet cap. Reasoning-token ratio above 80% of a step is a trip. Tool loops above six are a trip. Retry cascades accumulate history; the guard compacts or aborts before the retry is more expensive than the task.

Trust is an allowlist, not a hope. Production fleets load managed and certified only. Reviewed is for research-swarm. Community is for a laptop. The review agent is a different harness from the work agent. That last sentence is Meridian Loop 4.2.0, enforced in the runtime.

Human-on-the-loop is risk-weighted. IAM, payments, prod schema cannot auto-merge. A lint fix at 03:00 should not wait for a human who is asleep. Deloitte’s 2026 note called this the operating model of the firms that will still be here. It is an advance on this market, not a slide.

One thousand agents is an operations problem: regions, replicas, on-call, a degraded eval pool in Tokyo. Ten thousand is a market: license seats, hosted hours, attribution close, and a publisher who wants to be paid. The projector on the desk exists so you feel that line before you cross it.