Operations · 2026-08-21 · 16 min
One thousand is an operations problem. Ten thousand is a market.
At ten agents you have a config. At a thousand you have a topology, a budget, and an on-call. At ten thousand you have licensing, hosting, and attribution — or you have a cancellation.
IDC's installed-base forecast is the one that should make infrastructure people sit up: 28 million agents in 2025, 2.2 billion by 2030, task volume at a 524% CAGR. Even if you discount the billions, the direction is not ambiguous. The constraint is not 'can we spawn another worker'. The constraint is coordination.
A five-agent mesh has ten channels. A ten-agent mesh has forty-five. Zylos measured coordination overhead above 30% of the token budget in typical hub-and-spoke, and organisations running five or more agents without per-agent attribution reported 35–50% cost overruns. Gartner’s 40% cancellation rate by 2027 is the same story told by procurement: unanticipated cost, complexity of scaling, unexpected risk. Deloitte's 2026 prediction is the constructive version — orchestration is the inflection, and doing it well is worth 15–30% extra market by 2030 ($45B versus $35B on their autonomous-agent series).
Helix Supervisor exists because coding harnesses will not save you here. A coding harness is a worker. A supervisor is a topology (we refuse mesh above 80 agents), a budget (session cap $0.75, fleet hourly cap), a trust allowlist, and a human-on-the-loop router that decides by blast radius, not by a global pause button. Hierarchical beats hub-and-spoke above ~800 workers because the hub becomes the incident.
Hosting is the unglamorous half. Firecracker for managed labs, gVisor for OSS, Wasm where it fits. Regions are not a latency trophy — they are a data-residency and a failure domain. Scale-to-zero on the SRE fleet is a feature. An eval swarm that sits at 81% CPU in Tokyo is a smell. The desk is built to show both in the same rail.
The 1,000–10,000 band is specifically a software-delivery and research number, not a customer-support number. Support fleets went through this in 2024–2025 and mostly became a single vendor's contact-center product. Delivery fleets will not, because the git forge, the review policy, and the license mix are the company's. That is why the market for hosting, licensing, and git of harnesses is a separate object from 'buy Copilot'.