Hey there 👋 I read Okta’s Agent SSO announcement twice this morning, once for the news and once for the catch, because there is always a catch. The news is that AI agents now get a real identity inside the same directory that governs your employees. The catch is one line most of the coverage skipped: it ships inside core Okta SSO plans at no extra charge. When the identity incumbent gives agent identity away at no charge, read it as a default being set for the whole market. If I had to bet, this is the week agent identity stopped being an architecture project and turned into a checkbox. We have spent a month watching the cloud providers each mint their own version of it. This week it reached the login screen your entire company already stares at every Monday, and that is a different kind of arrival. The honest limit, said up front: shipping a standard is not the same as your apps supporting it. Your real gate is coverage, which tool in your stack speaks the protocol yet. More on that below. Four stories, then the bottom line. The thread: agents got normal this week, and every launch was about making them safe to run.
The Big ThingOkta ships Agent SSO to GAOkta moved Agent SSO to general availability, folding its open Cross App Access standard (XAA) into the core Okta identity product that already runs for more than 20,000 customers. In plain terms: an AI agent registers as a first-class identity in Universal Directory, gets assigned policies, and receives short-lived, scoped tokens instead of a hard-coded API key buried in someone’s script. Same lifecycle, same access reviews, same offboarding you already run for people. The reason this lands harder than the cloud-provider versions is distribution. Agent identity is included in core Okta SSO plans at no extra cost, and the early-adopter roster is the software your team already opens: Anthropic, Asana, Atlassian, Cloudflare, Cursor, Datadog, Figma, Linear, Slack, Zoom. When the identity you already pay for treats an agent like a badge-carrying contractor, the greenfield question of how to mint agent identities mostly evaporates. What replaces it is a coverage problem, which is more boring and more real. Cross App Access only helps for the apps that implement it, and today that is a real list, not a universal one. So the design-review question shifts from how do we build this to which of our SaaS tools speak XAA yet, and what do we do for the ones that don’t. For the holdouts you are back to scoped service accounts and a rotation policy, same as last week. Grade the interop claims carefully. This is Okta’s framing of Okta’s launch, so read the partner support as vendor-stated until your own tenant proves it. The standard is open and the adopter list is public, which is more than most agent-identity pitches can show, but supported on a slide and wired up in your tenant are two different Tuesdays. Ship it? If you are on Okta, pilot XAA on one real agent workflow this week, ideally an agent that already touches two or three of the launch-partner apps. Inventory every agent quietly holding a long-lived key and treat that list as your migration backlog. If you are not on Okta, read the spec anyway, because it is the clearest signal yet that agent identity is going to ship inside the tools you own, not as a box you buy. Sources: Okta newsroom, Okta brings first-class identity to AI agents with Agent SSO (Aug 26, 2026); TechNode coverage (Aug 26, 2026).
Tour de Headlines💬 Salesforce puts AI coding in a shared Slack channel. Salesforce launched Slack Code on Aug 24: dedicated code channels where a team tags a coding agent and watches it work in the open. Inside a channel you get tabs for the conversation, the agent’s plan, line-by-line diffs, and a live preview of the running output, so a PM or designer can follow the build without reading a terminal. It is agent-neutral, wiring in Claude Code, Cognition’s Devin, GitHub Copilot, ChatGPT, and Vercel’s agents, and anyone in the channel can pause, redirect, or stop a run mid-task. Human sign-off is still required before an agent pushes to production. The price is the tell: Slack Code is free on every Slack plan, you just bring your own agent subscriptions. Upside, agent coding becomes an auditable team activity instead of one person’s private tab. Risk, more hands near the repo, so keep branch protection, CI, and required review exactly where they are. Turn it on for a staging repo first. Sources: Salesforce newsroom, Slack Code launch (Aug 24, 2026). 📊 The data says agents are daily drivers now. Temporal’s 2026 State of Development: AI Agents report (554 engineers and eng leaders across the US and UK/EMEA) puts a number on what your standup already told you: 80.8% now use AI agents daily or more, up from 47.3% a year ago. Adoption is a closed question. Reliability is not. The top blocker to using agents more was tracking state at 35.7%, ahead of debugging and managing token or compute cost. That ranking is the useful part, because it says the wall teams hit is not model quality, it is the unglamorous execution layer: keeping an agent’s state coherent across retries, failures, and long-running work. Teams the report flags as successful were 1.5x more likely to run agents daily, which reads less like a magic model and more like they built the plumbing to trust one. Fair caveat: Temporal sells durable execution, so the framing points at the problem their product solves. The adoption numbers are worth carrying into your next planning meeting anyway. Sources: Temporal, 2026 State of Development: AI Agents (survey Apr 29 to May 25, 2026). 🚦 Google Cloud calls agent security the gate, not the trim. Google Cloud published guidance on Aug 24, tied to its State of AI Infrastructure work, that lands on the same nerve as the Okta launch from the platform side. Its argument: agent security is the top gating issue for scaling autonomous workflows, and the fix is platform-level governance, task-level provenance, short-lived permissions, and human-in-the-loop checkpoints, not agents bolted onto legacy access and logging. There is no product to deploy, so grade it as a point of view, not a release. It is a useful one, because it names the exact controls to ask your vendors about: can you trace a single agent task end to end, can you scope a permission to one action and expire it, and where does a human have to sign. If you are standing up an agent that writes, pays, or provisions, instrument those three before you widen its autonomy. This is the report’s recommendation, not a benchmark, so treat it as a checklist to pressure-test rather than gospel. Sources: Google Cloud blog, State of AI Infrastructure guidance (Aug 24, 2026).
Sponsor Your agents get a verified identity. Do your reps? RapportScore measures how your reps communicate on real calls, then coaches the behavior that builds trust. Deterministic signals you can coach against, not vibes. See where your team stands. See your team’s score → |
Tool of the Day🔎 KeenableA live-web search backend built for agents, with an MCP endpoint, so your agent pulls fast, cited web results without you running a crawler or index. Keenable came out of stealth on Aug 25 with a $26M seed led by Accel (Conviction and angels from Amazon, ClickHouse, Databricks, Google, and Snowflake also in). The founders are Andrey Styskin, who ran search at Yandex, and Matthias Petri. The product is an independent 100 billion-plus document index with a low-latency Search API, page-content retrieval, and an MCP interface that has a free keyless tier for prototyping. The use cases they name, lead enrichment, market mapping, and pricing monitoring, are exactly the high-frequency lookups agents choke on with a general search tool. Two honest caveats: it is a days-old seed-stage company, so treat any production use as a pilot, and the index is theirs, not something you can audit from the outside. If web retrieval is your agent’s weak point, a drop-in MCP endpoint is the cheapest way to test whether a purpose-built index beats what you run today. Read the launch details →
Worth a Click- Strip your agent logs before you publish them. A new arXiv paper decoded 315,000 hidden reasoning blocks from public agent traces and recovered 62 API keys, 33 passwords, and 7 private keys. Vendors say the main extraction path is now mitigated, but the habit stands: scrub reasoning and opaque fields from any trace you share.
- Two agent standards moved under one roof. Google’s A2A protocol joined the Linux Foundation-directed Agentic AI Foundation, sitting next to Anthropic’s MCP, with 250-plus members including AWS, Google, Microsoft, and OpenAI. Fewer competing standards means less glue code later.
- The adoption number has a shadow. Set against Temporal’s 80% daily-use figure, one widely cited report finds roughly 99% of companies plan to run agents in production and only about 10% have. The distance between pilot and production is architecture and governance, not appetite.
The Bottom LinePut the week on one screen and it splits in two. On one side, 80% of engineers drive an agent every single day. On the other, only about one company in ten runs an agent in production. The daily driver is a copilot sitting next to a human who signs for everything it does. The production system is the thing still waiting on identity, durable state, and a governance story a security team will initial. Read this week’s launches in that light and they rhyme: Okta made agents first-class identities in the login you already own, Salesforce put agent coding where the whole team can watch it, Google Cloud wrote down which controls gate the rollout. None of this is a smarter model. It is the boring, load-bearing plumbing that decides whether the agent you use daily is one you would ever let sign for money or data. Everyone has an agent. Almost nobody has one they would trust unsupervised. Closing that gap, not the next benchmark, is the whole 2026 market. More tomorrow,
— Ron |
