2026–2030: a desk forecast

Houses do not agree on the size of 'the agent market' because they do not agree on the noun. Grand View and the narrower MarketsandMarkets series put 2026 around $11B and 2030 around $50–53B at ~46% CAGR. The broader agentic series (MarketsandMarkets August 2026) puts 2026 at $19.33B and 2033 at $205.88B. Deloitte's autonomous-agent cut is smaller ($8.5B → $35B) and explicitly says orchestration quality is worth 15–30% extra. Presenc's composite — $12.06B in 2026 to $53.2B in 2030 — is what this desk uses for the agent layer. Our own split of that layer is the useful part.

2026, this year. Mixed fleets become the default in software delivery. Agent Skills is the rare cross-vendor standard. Five distribution surfaces matter: Microsoft Marketplace, Salesforce Agentforce, Gemini Enterprise, ChatGPT Workspace (not a public catalog), and the open Skills directory. Gartner’s cancellation clock is already ticking on projects that have a demo and no supervisor. Human-on-the-loop is the operating model of the firms that will still be here in 2028. Coding agents remain the largest role by spend; vertical agents are the fastest (62.7% CAGR).

2027. The cancellation year. Forty percent of agentic projects die, per Gartner, of cost, scale, or risk. Survivors have three objects: a supervisor, an allowlist, and a trace layer. Dual-license harnesses standardise. Insurance questionnaires start asking for attribution. Malicious-skill scanners become a line item, not a blog post. Mesh topologies are professionally embarrassing above 80 agents.

2028. Compounding. Kaiso's enterprise series puts 2028–2032 as the years departmental agents become cross-functional systems. Multi-agent already holds 53.3% of 2026 agentic revenue; it will be the only architecture anyone admits to. Fifteen percent of day-to-day business decisions autonomous, per Gartner. Hosted harness runtime is a real cloud category — our series puts hosted at $7.4B in 2028. Git forges (GitHub, GitLab) are either a harness vendor or a registry, sometimes both.

2029–2030. The IP layer. License liability is priced. Attribution is a close process, not a dashboard. Orchestration ($17.2B on our series) plus hosted ($21.4B) plus skills ($10.6B) is larger than the residual 'agents' narrative. IDC's 2.2 billion agents is only plausible if most of them are short-lived workers under a supervisor, not pets. BCG's $200B tech-services expansion is the services shadow of this stack: the firms that cannot hire harness engineers will rent them. The labs still matter. They no longer set the bill of materials.

What would make this forecast wrong: a single lab ships a supervisor so good that mixed fleets die; a standards war fragments Skills; regulation treats agent output as unlicensable. We do not underwrite those. We underwrite the boring path — mixed fleets, portable capabilities, paid orchestration, argued-over percents.