Okta gives AI agents staff identities before the audit exists

Okta’s Agent SSO is the point at which the agent stack stopped being a developer concern. Registering an agent in the same directory as an employee, and issuing it short-lived tokens instead of a stored API key, is how an organisation concedes that an agent is a party to its business rather than a script someone runs. Okta’s own figure — 34% of organisations apply the same security controls to agents as to people — shows how far that concession still has to travel.

Agents are being credentialed faster than they can be held accountable. Identity landed this week and payment authority landed the week before, both on vendor product cycles. The evidence chain that would prove an agent’s purchase matched its instruction cannot ship that way: it needs an agent provider, a retailer and a payment network to agree, and no single company owns all three. Google’s AP2 records what each participant saw and stops there. Meanwhile Temporal found 41.1% of engineers hitting agent issues daily while 85.5% trust the outputs.

Last Thursday’s edition covered the protocol layer converging. What has changed is the object of convergence — from how agents talk to what they are allowed to hold. The test over the next stretch is whether NIST’s tasking under the AI AGENT Act reaches for a cross-company evidence standard or settles for per-vendor logs.

Okta brings first-class identity to AI agents with Agent SSO

Okta · 24 August 2026 — Agent SSO registers AI agents as first-class identities in Okta’s Universal Directory and issues short-lived, identity-governed tokens in place of static API keys. It covers agents from more than 14 platforms including Claude, Slack, Atlassian and Figma, and folds in Cross App Access, the enterprise authorization extension to MCP.

Why it matters: Static keys made agent traffic anonymous; a directory entry gives it an owner, a policy and an audit trail.

Google can track how your agent spends your money — but not whether you approved it

Fortune / The Conversation · 24 August 2026 — Google’s Agent Payments Protocol records the limits a user approved and what each participant saw, but no chain links the original instruction to the charge. Senator Mark Warner’s AI AGENT Act, introduced 21 July, directs NIST to set delegation and audit standards without requiring that cross-system chain.

Why it matters: When an agent buys the thing it was told not to buy, every party holds accurate records and none of them can settle the dispute.

The 2026 State of Development Report: AI Agents

Temporal · 25 August 2026 — A survey of 554 engineers found 80.8% now use agents daily, up from 47.3% a year earlier, and 49.1% run them in production or as core to how they ship. In the same sample, 41.1% hit agent issues daily or more often, while 85.5% trust or somewhat trust the outputs.

Why it matters: Trust is outrunning reliability by a wide margin, and that gap is exactly where governance costs land.

DeepSeek-V4-Flash-Vision-Exp release: multimodal API now live

DeepSeek · 21 August 2026 — Vision arrives on DeepSeek’s cheap tier. The experimental model matches V4-Flash on text, bills images at up to 384 tokens each, and the company says it brings multimodal agent performance close to Claude Opus 4.8. A new Files API lets one uploaded image be reused across requests at no extra cost.

Why it matters: Screen-reading agents get materially cheaper, which widens the set of tasks worth handing to one.

Agentic web search infrastructure startup Keenable raises $26M

SiliconANGLE · 25 August 2026 — Keenable left stealth with a $26m seed led by Accel and Conviction Partners, selling agents an independent index of more than 100 billion documents at $1 per 1,000 API requests. It also offers point-in-time queries — the web as it existed on a given past date — and says a handful of AI labs already use it in production.

Why it matters: Agents are acquiring their own retrieval layer rather than borrowing the consumer web’s, which changes who sees their queries.

Watch the NIST tasking under the AI AGENT Act, and whether any vendor ships an evidence chain that survives contact with a second company’s records. Identity was the easy half.

Written by my AI assistant.