Desk · 2026–2030
The money moves up the stack. The risk is a missing primary key.
Houses do not agree on the noun. This desk splits the layer: agents, hosted runtimes, skills, orchestration — and writes the operating consequences.
Desk series · USD billions
What the houses said
| House | 2026 | 2030 | CAGR | Note |
|---|---|---|---|---|
| Grand View Research | $10.9B | $50.31B | 45.8% | AI agents, 2025–2030. The cleanest single series for the narrow agent market. |
| MarketsandMarkets (agents) | $11.47B | $52.62B | 46.3% | Derived 2026 from a 2025 base of $7.84B. Vertical agents at 62.7% CAGR. |
| MarketsandMarkets (agentic) | $19.33B | — | 40.2% | $205.88B by 2033. Broader than 'agents' — includes orchestration and infrastructure. |
| Presenc / desk composite | $12.06B | $53.2B | 44.9% | The figure this desk uses for the agent layer unless a client asks for a named house. |
| Deloitte (autonomous agents) | $8.5B | $35B | 42.3% | Orchestration done well adds 15–30%, to ~$45B. The extra is the product we sell. |
| IDC (installed base) | — | — | 524% | 28M agents in 2025 → 2.2B by 2030. Task volume 415 trillion. Orchestration is the fastest-growing spend layer. |
Houses · 2026 vs 2030 · $B
Only series with a printed year. IDC is an installed-base CAGR, not a dollar stack — it stays in the table.
Floor · kind and pay
Derived from the 519-slip catalog. Cited scaffolding premium sits beside it so the two objects stay separate.
Scaffolding premium · Δ pt
Briefs
The harness is the product
Models commoditise. The software shell that turns them into agents does not. 2026 is the year the industry started saying this out loud.
Skills, MCP, and the capability market
The unit of exchange is no longer an agent. It is a capability with a permission manifest. The graph is already too large to browse.
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.
How much percent of each harness
Token share is a cost metric. Accepted-work share is a productivity metric. Shapley-lite is a contribution metric. License liability is a legal metric. They will not agree. That is the point.
Licensing, git, and the coming IP layer
A Skill is a work. A harness is a work. A mix is a process. Someone will want to know who owns each, who may run each, and which SHA did.
2026–2030: a desk forecast
Five years. Four layers. One cancellation cliff. The money moves up the stack: from models, to harnesses, to orchestration, to attribution.
Why forty percent die
Gartner’s 2027 cancellation rate is not a mystery. It is three missing primary keys: a budget with a breaker, a topology with a supervisor, and a mix with a number.
HMX: a mix index
Industry production-token share, Q3 2026. Not a ranking of quality. A ranking of who is actually on the bill.
Five hundred and nineteen slips
The open index is a pointer file, not a warehouse. After the August 2026 landscape fill: 116 harnesses, 128 skills, 144 advances, 131 MCP. GitHub is how most of them arrive. Open is how most of them price.
Verified is saturating. Pro is not.
29 August 2026. BenchLM Verified top three within a point. Scale live Pro and Terminal-Bench 3.0 still have air. Do not subtract them.
Method
Named houses (Grand View, MarketsandMarkets, Deloitte, IDC, Presenc, Skillful.sh, Zylos, Gartner via secondary). Desk composite $12.06B (2026) → $53.2B (2030). Hosted / skills / orchestration splits are ours. HMX is production-token share, not seats.
Eval
Pass@1 holding the model constant, dollars per resolved, mix disagreement, scale overhead — live on the benchmarks board.
Open benchmarks