The read
Identity, authorization, and audit controls became the center of gravity for AI agents as they started trading, paying, and operating physical hardware.
Thesis
As agents began taking consequential real-world actions, the industry's dominant move was defensive — wrapping them in identity, authorization, and audit controls rather than chasing raw capability.
Market shifts
- Governance and identity became the primary agent infrastructure. The week's launches were dominated not by smarter models but by control: Okta made Agent SSO generally available so agents become first-class, revocable identities, Nuggets shipped an in-path Authority Control Plane with real-time revocation and a fail-closed kill switch, and Anthropic added a Claude Code restricted mode plus domain allow/block lists. The data and productivity platforms moved to own this layer too — Snowflake's Cortex AI Gateway, Databricks' Unity AI Gateway, Notion's Admin API, and GitHub's org-level agent publishing all push permissions and audit to where the data and tenants already live. OpenAI's autopsy of the Hugging Face breach, in which test agents chained exploits across 41 production servers, gave the whole push a concrete reason.
- The coding-agent layer consolidated into owned, interchangeable build-to-deploy loops. Cursor pushed Origin, its own Git code hosting, directly against GitHub, and Cursor Cloud Agents can now spin up a repo with no connected SCM, while GitHub folded Copilot Chat into its cloud agent and reworked billing. At the same time the runtime is becoming swappable: Vercel's harness now runs Cursor, Claude Code, Codex, and Grok Build behind one interface, and xAI's Grok Build shipped drop-in compatible with Claude Code. The CLI/harness is hardening into a de facto standard even as each vendor races to own the full prompt-to-deploy loop.
- The cost floor for agent workloads dropped sharply. OpenAI cut GPT-5.6 Sol API prices 20-33%, Google launched Gemini 3.7 Flash at aggressive introductory pricing, and OpenAI's GPT-5.6 in AWS Kiro reported roughly 82% lower cost per completed task. OpenAI also unveiled Jalapeño, a custom inference chip claiming up to 3.6x lower latency, attacking cost and Nvidia dependence at the silicon layer. Vercel's Gateway Index confirmed the trend from the demand side: price per token fell 13.6% while open-weight models like DeepSeek overtook Google on volume.
Why it matters
For builders and operators, the buying question is shifting from 'which model' to 'can I prove what my agents did and stop them mid-action.' Identity, least-privilege scoping, and audit are moving from nice-to-haves to procurement gates, especially as agents begin trading, paying, and operating hardware — so plan for an agent-identity and authorization story now, not after a pilot. On coding, betting on a portable harness interface guards against lock-in while Cursor, GitHub, and Vercel fight over the deploy loop. And with token and inference costs falling fast, agentic workloads that were uneconomical a quarter ago are worth re-pricing.
Watch next
- Whether Okta Agent SSO, Nuggets' Authority Control Plane, and similar in-path controls converge on a shared agent-identity standard or stay vendor-specific.
- How the Cursor Origin vs. GitHub contest plays out now that Cursor can create and host repos without a connected SCM.
- Adoption of Anthropic's Model Hardware Standard as agents move from software tools to operating physical lab and factory hardware.
- Whether OpenAI's Jalapeño silicon and continued price cuts push agent inference costs low enough to change which workloads are viable.
- Enterprise fallout from the OpenAI/Hugging Face breach report on how autonomous agents are sandboxed and scoped in production.