Somewhere in your repo is an environment variable named something like AGENT_API_KEY. It holds a long-lived credential your agent uses to act on a customer’s behalf. It belongs to no person, expires never, and gets rotated when someone remembers. Until this month that was normal engineering. In August it became legacy.

Two announcements did it. On August 24, Okta shipped Agent SSO into general availability, included in its core SSO plans at no additional cost. An agent gets an entry in the directory like an employee, a policy that governs it, and a short-lived token when it acts inside another application on a user’s behalf. Security vendors charge for differentiators and bundle table stakes, so the pricing is the real announcement here: Okta has decided agent identity is plumbing that every customer will need.

A week before that, Google moved its Agent2Agent protocol into the Agentic AI Foundation, the Linux Foundation body that has stewarded MCP since Anthropic donated it in December. The two protocols that matter, MCP for connecting agents to tools and A2A for connecting agents to each other, now sit under one roof. The foundation’s platinum members include Anthropic, OpenAI, Google, Microsoft, AWS, and Block. Okta is a member as well. The companies competing hardest at the model layer are jointly funding the layer below it.

The fragmentation everyone priced in never arrived

Through 2025, waiting was a defensible engineering position. The fear had a familiar shape: commit to a protocol, watch a rival standard win, rewrite your integration layer in eighteen months. Teams filed agent support under “revisit when the standards settle” and moved on.

The adoption numbers stopped supporting that fear some time ago. MCP passed 10,000 published servers with client support in ChatGPT, Claude, Gemini, Microsoft Copilot, Cursor, and VS Code. A2A crossed 150 supporting organisations, shipped a stable 1.0, and landed in Google Cloud, Azure, and AWS. August removed the last plank. When the sponsors of rival ecosystems pay for the same foundation, the war you were waiting out has been called off.

Back in March I wrote that the agent OS war had started and argued for building to abstractions, because the platform layer was undecided. The platform layer is still undecided; the protocol layer underneath it settled, and hedges built for the first question now tax teams on the second.

Why the engineers on product teams carry this one

Agent identity sounds like a security team’s problem. Watch where it lands. Okta’s own research found only 34% of organisations apply the same security controls to AI agents as they do to human workers. That gap was tolerable while nobody could name the fix. Agent SSO builds on Cross App Access, an open standard rather than an Okta-only mechanism, and it is now bundled into a product thousands of enterprises already run.

Enterprise security reviews follow available controls with a short lag. The questionnaire that asks about SSO and SCIM today will ask how agents authenticate to your product, and how your product’s agents authenticate to everything else. Once the control is cheap and standard, a missing answer looks like a finding. The engineers who design your auth flows will do this work either ahead of that questionnaire or in the middle of a stalled deal.

Three moves before the quarter closes

First, if your product has an API, give it an MCP server built against the spec itself; skip the compatibility shim. Agents are becoming a real share of how products get operated, and MCP is the surface they arrive through. Treat it the way you treated your public API a decade ago.

Second, inventory agent credentials, in both directions. Count the static keys agents hold against your product and the ones your own agents hold against everything else. Replace them with short-lived, policy-scoped, delegated access. Cross App Access gives you a standard to design against even where Okta isn’t in the picture.

Third, retire the abstraction hedge if you built one. A translation layer between your code and “whichever protocol wins” was reasonable insurance in 2025. You are still paying the premium on a risk that no longer exists.

The waiting position had real logic, and it expired in August. Somewhere in a backlog near you sits a ticket that still says “revisit when the standards settle”. Close it with us.

Sources