OpenAI just dropped the Agents API into public beta, and the pitch is seductive: build and launch cloud agents with a managed service powered by the Codex harness. Translation? The company is taking the hard, fiddly parts of agent infrastructure — session state, retries, context compaction, multi-step recovery — and making them someone else's problem. Yours, if you're a developer. Theirs, if you're a competitor.
Here's the sharpest edge: there are no additional fees for using the Agents API. You pay for what you use — meaning tokens. OpenAI isn't monetizing the orchestration layer. It's giving it away to lock in the metered layer underneath. That's the classic platform play: commoditize your complement. And it works. Every agent built on this API is a token-consumption engine.
But here's the twist: Google got there first. Back in July 2026, Google shipped the Genkit Agents API in preview for TypeScript and Go, with a signature feature called 'detached turns' — agents that keep working after the client disconnects. Read that next to OpenAI's 'long-running sessions' and the pattern snaps into focus. Both companies converged on the same insight within about eight weeks: the agent problem is no longer intelligence. It's persistence. The agent that dies when you close your laptop is the agent nobody ships to production. Google's take is open-source, multi-language, framework-shaped. OpenAI's is managed, hosted, API-shaped. One asks you to adopt a library. The other asks you to trust a platform.
So who wins? Application developers and small teams, immediately. The single biggest barrier to shipping an agent was never the model — it was the unglamorous plumbing. Bundling orchestration, compaction, and recovery into one call collapses months of engineering into an afternoon. OpenAI wins too, obviously. And the MCP ecosystem gets a boost: OpenAI supporting MCP servers means Anthropic's de facto tool-calling standard gains legitimacy from its fiercest competitor. Tool vendors who built to MCP now have a second major distribution channel.
Who loses? The agent-orchestration startup category. This is structural inference, not reported fact — but the logic is unavoidable. If a managed harness with orchestration, subagents, and compaction is free at the platform level, the standalone orchestration layer is being squeezed from above by the company that also owns the model. Same squeeze that hit vector-database startups when managed embeddings arrived. Third-party sandbox providers? Also feeling the heat. The announcement explicitly includes 'hosted sandboxes' and 'pluggable' execution environments. 'Pluggable' is generous framing; 'hosted' is competition. And Google, partially: Genkit has the head start and the open-source credential, but OpenAI has default distribution. Being first with a preview matters far less than being default with a GA launch.
The tension nobody's talking about yet: the 'free' that isn't. 'No additional fees' is a pricing philosophy, not a gift. It resets developer expectations so that any competitor charging for orchestration looks like a toll booth. Expect this to be revisited once switching costs are established — the no-extra-fee era is a land-grab, and land-grabs end. There's also the lock-in question: OpenAI supports MCP (interoperable) but manages sessions, state, and recovery (proprietary). Tool-level portability, infrastructure-level lock-in. That's a trade developers will make happily — until they can't leave.
My prediction? Trends suggest this is the opening shot in a platform war over who owns the agent runtime. Google will respond by pushing Genkit toward GA and leaning harder into open-source as a differentiator. The orchestration startups will either pivot to vertical-specific agents or get acquired. And within a year, 'no additional fees' might quietly become 'no additional fees for the first million calls.' The land-grab always ends. The question is whether you'll notice when it does.