HarnessmarketEnter the desk
/HMX Claude Code 29.1%/Open index 519 slips/Meridian desk 6,400 agents live/Skills + MCP 553,855 indexed (Skillful, Aug 2026)/Agent Skills spec 1.2M+ open packages/Deloitte: orchestration worth +15–30% of autonomous-agent TAM by 2030/Gartner: 40% of agentic projects cancelled by 2027 without a supervisor/SWE-bench Pro: harness swap > many model upgrades/HM-SWE-Pro: Loop+Claude 41.2 pass@1 · chat 18.2 · same model/Lewis 2608.26218: F2PF 28→49 under a tighter harness, same model/ACES: Skill Lift 0.21 · 947 paired cases · scan vs live ρ=0.14/Tokenomics: code review 59.4% of ChatDev tokens/HMX Claude Code 29.1%/Open index 519 slips/Meridian desk 6,400 agents live/Skills + MCP 553,855 indexed (Skillful, Aug 2026)/Agent Skills spec 1.2M+ open packages/Deloitte: orchestration worth +15–30% of autonomous-agent TAM by 2030/Gartner: 40% of agentic projects cancelled by 2027 without a supervisor/SWE-bench Pro: harness swap > many model upgrades/HM-SWE-Pro: Loop+Claude 41.2 pass@1 · chat 18.2 · same model/Lewis 2608.26218: F2PF 28→49 under a tighter harness, same model/ACES: Skill Lift 0.21 · 947 paired cases · scan vs live ρ=0.14/Tokenomics: code review 59.4% of ChatDev tokens

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.

Eval desk

Desk series · USD billions

0.0265177102202520262027202820292030
agentshostedskillsorchestration

What the houses said

House20262030CAGRNote
Grand View Research$10.9B$50.31B45.8%AI agents, 2025–2030. The cleanest single series for the narrow agent market.
MarketsandMarkets (agents)$11.47B$52.62B46.3%Derived 2026 from a 2025 base of $7.84B. Vertical agents at 62.7% CAGR.
MarketsandMarkets (agentic)$19.33B40.2%$205.88B by 2033. Broader than 'agents' — includes orchestration and infrastructure.
Presenc / desk composite$12.06B$53.2B44.9%The figure this desk uses for the agent layer unless a client asks for a named house.
Deloitte (autonomous agents)$8.5B$35B42.3%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.

0.013274053Grand View ResearchM&MM&M ag.PresencDeloitte
20262030

Floor · kind and pay

Derived from the 519-slip catalog. Cited scaffolding premium sits beside it so the two objects stay separate.

advance14427.7%
mcp13125.2%
skill12824.7%
harness11622.4%
open32462.4%
seat9217.7%
usage7314.1%
enterprise285.4%
one-time20.4%

Scaffolding premium · Δ pt

Minimal → Claude Code
36.0 pt
Stanford IRIS Meta-Harness
33.4 pt
Claw-SWE-Bench max gap
27.4 pt
Codex native → Cursor
25.7 pt
Qwen3-32B scaffolds
20.0 pt
LangChain harness swap
13.7 pt
Auggie
5.9 pt
Cursor
4.3 pt
Claude Code
3.9 pt

Briefs

Thesis2026-08-24 · 14m

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.

Market2026-08-29 · 12m

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.

Operations2026-08-21 · 16m

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.

Attribution2026-08-29 · 13m

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.

Market2026-08-12 · 11m

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.

Forecast2026-08-29 · 18m

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.

Operations2026-07-30 · 9m

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.

Index2026-08-29 · 6m

HMX: a mix index

Industry production-token share, Q3 2026. Not a ranking of quality. A ranking of who is actually on the bill.

Index2026-08-30 · 8m

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.

Eval2026-08-29 · 7m

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