Field Guide
AI agents: orchestration platforms
06
Market and trends
4 min read · Chapter 6 of 10
Who holds the power
Power follows the model, the data and the identity layer, and an independent platform holds none of them.
- Labs set prices, caching rules, rate limits, retention and retirement dates, and they moved up the stack in 2025 and 2026. OpenAI reportedly launched an enterprise agent platform in February 2026, and bought an experimentation startup (about $1.1B in stock) and an open-source eval and red-teaming tool. Anthropic hired an eval startup's team in August 2025; that startup's platform went offline four weeks later. Google began billing for agent runtime, sessions and memory in December 2025. Labs withdraw layers too, like OpenAI's hosted Evals. I found no similar moves by Meta, Mistral or xAI; Meta's tools are open weights and published security guidance.
- Labs and clouds co-govern the protocols: the Agentic AI Foundation's platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
- Clouds sell governance at fractions of a cent, already run the enterprise's identity and network, and set their own model dates.
- Systems of record own the data and workflow: ServiceNow's AI passed $1B in annual contract value (July 2026), and it paid $2.85B for an agent platform in 2025.
- Security incumbents bought AI-security startups in 2025: Palo Alto Networks, Check Point and SentinelOne, reportedly for about $250M to $700M each.
- Customers have routing (sending each request to the cheapest model that can handle it; on one benchmark, research routers cut cost more than 85% and kept 95% of quality), falling prices and cloud commitments that pay for any marketplace model.
What locks a customer in: eval datasets and graders, traces, memory stores, connectors, one lab's caching economics, fine-tunes, commitments. Gartner reckons only about 130 of thousands of self-described agentic vendors are real. This next part is my own conclusion rather than a reported fact. Labs and clouds are absorbing the control layer from above and systems of record the workflow from below, so an independent platform is left with neutrality, meaning state, evals and traces that move across labs.
Tailwinds and headwinds
These are the forces I'd expect everyone in this industry to watch over the next couple of years. Each one helps some players and hurts others; the clouds end up on the winning side of most of them, and the independents win mainly where neutrality across labs is the point.
Tailwinds:
| What's changing | Who gains | Who loses |
|---|---|---|
| Large companies are raising AI budgets again: in KPMG's 2026 survey, planned AI investment was up about a third on a year earlier. | The clouds, where most model spend lands; the labs; systems of record selling agents (Salesforce, ServiceNow, Workday); agent builders (Sierra, Decagon) | Software budgets without AI, which fund the new spend |
| Agents get paid for from labor budgets, and Gartner expects agents to resolve 80% of common customer service issues without a person by 2029. | Agent builders priced per outcome (Sierra, Decagon); systems of record with service agents (Salesforce, ServiceNow) | Vendors priced per seat, when the seats are the jobs being cut; probably outsourced contact centers |
| The price of a given level of model performance keeps falling, by about 50 times a year at the median in Epoch AI's data. | Clouds and platforms billed on volume; agent builders priced per outcome, whose margin widens; buyers | The labs' revenue per token; tools whose main pitch is cutting model cost, such as routers and caching layers |
| Most large companies now use models from several labs, and in an a16z survey of large-company CIOs 81% used three or more model families. | The clouds, which resell many labs; neutral platforms that hold state, evals and traces across labs (LangChain, Temporal, Braintrust, Arize, Datadog) | Any lab's own agent platform sold as lock-in; buyers who built on one lab's stateful API |
| Agents run longer and in teams, and in KPMG's survey the share of large companies running multi-agent systems roughly quadrupled in one quarter. | Durable-execution engines (Temporal); data platforms that hold agent state (Databricks, MongoDB, Snowflake); the labs' and clouds' managed runtimes | Thin wrappers that only chain prompts; buyers on consumption pricing, since cost grows faster than turns |
Headwinds:
| What's changing | Who gains | Who loses |
|---|---|---|
| The labs now sell agent platforms and finished tools of their own, competing with the companies built on their models. | OpenAI, Anthropic and Google (I found no similar move by Meta, Mistral or xAI) | Independent orchestration and eval vendors (LangChain, Braintrust, Arize); seat-priced systems of record (Salesforce, Workday, SAP, ServiceNow); agent builders whose product a lab copies |
| Demand for computing power outran supply in 2026, bringing rate limits, outages and capacity sold on terms of up to three years. | The clouds and labs, which gain pricing power; the biggest buyers, who can reserve capacity; platforms that fail over between labs | Startups and scale-ups on self-serve tiers; agents tied to one lab |
| Clouds and software suites now include agent governance in bills buyers already pay, such as Microsoft 365 E7 at $99 per user a month. | Microsoft, AWS, Google Cloud | Standalone gateway, policy, registry and eval vendors; Okta, where Microsoft Entra comes in the bundle |
| The apps that hold company data limit outside AI access while selling agents of their own, as Slack's 2025 API terms did. | Salesforce, ServiceNow, SAP, Workday | Glean and other agents that index other apps' data; independent platforms that need connectors; buyers who want one agent across apps |
| Many buyers still can't show a return, and Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027. | Vendors priced per outcome; incumbents buyers already trust; clouds selling pay-as-you-go | Licence-priced agent platforms without measured results; vendors relabeling old products as agents |
The one thing I'd watch is whether companies that use several labs build their next agents on a lab's own platform or on a neutral one, because that decides whether the independents keep a reason to exist.