This week, the story isn't a single agent launch — it's how fast the ecosystem around agents is institutionalizing, especially on security and identity, while adoption keeps compounding in specific verticals.

This week's developments

Cisco has rolled out its "MyAgent" personalized agentic AI assistants to all 90,000 employees globally, and says agentic interactions across the company are up roughly 350% quarter over quarter. It's one of the largest confirmed enterprise-wide personal-agent deployments disclosed to date, and a real signal that usage compounds fast once a company commits at scale. (Source: Cisco Blogs)

Three major security vendors moved to formalize agent governance in the same week. CrowdStrike launched an AI Partner Specialization with a "Verified Agent" certification for what it calls the "agentic enterprise," building on its Falcon platform with Accenture, Anthropic, CoreWeave and WWT as named partners. Proofpoint shipped a SOC Analyst Agent built on OpenAI's Daybreak models, targeting general availability by the end of Q3 2026. Tenable launched CyberAgents Exchange AI Inspector, pairing OpenAI's GPT cyber models with human researcher review to vet third-party agents before enterprises deploy them. Taken together, they read as the security industry deciding agent vetting is now a permanent product category, not a stopgap. (Sources: CrowdStrike, AI Agent Store)

That governance push has real justification behind it: Manifold Security disclosed "GitSpawn," a set of 8 vulnerabilities spanning 7 AI coding agents — including Claude Code, Codex, Cursor, Goose, Hermes Agent, Qwen Code and Grok Build — that exploit Git's core.fsmonitor setting to trigger code execution via malicious .git configs before a user ever approves anything. Four of the eight flaws were still unpatched at disclosure. (Source: The Hacker News)

Separately, OpenAI's GPT-6 Astra became the first model to cross OpenAI's own "Critical" cybersecurity threshold, assessed as capable of developing exploits against hardened systems. Access is now gated to OpenAI's Trusted Access and Daybreak programs rather than shipped broadly — the first time a frontier lab has had to formally restrict a release on these grounds. (Source: MarkTechPost)

On the money side: Okta's Q2 FY2027 results showed $805M in revenue (up 11% YoY) and $116M in GAAP net income, and the company raised full-year guidance to $3.216–3.226B, explicitly citing demand for AI agent identity and access management. It's one of the clearer signs that enterprises are already putting real budget behind securing what agents are allowed to touch. (Source: Okta newsroom)

Vertical automation kept expanding too. Cashfree's "Relay AI" Super Agent moved from beta to general availability for SMB payment operations in India, with the company targeting recovery of ₹20,000 crore in stalled merchant transactions and claiming it cuts roughly 45 minutes a day from the ~60 hours a week SMBs spend on payment-related work. Optimizely introduced "Virtual Teammates" — role-specific agents for marketing functions like Chief of Staff, SEO Analyst and CRO Manager — backed by its own survey of 2,000+ B2B marketing leaders, 81% of whom say they're juggling two or more disconnected AI tools weekly. (Sources: Business Standard, PR Newswire)

Funding kept flowing into agent infrastructure and vertical plays: AIR Security raised $50M across two seed rounds (led by Sequoia and Greenoaks) to vet the AI agent software supply chain, and says it's already filtering roughly 27% of agent add-ons as unsafe across 20+ customers, mostly in finance and pharma. The same stretch also saw Aslan raise $20.8M for AI agents in national-security investigations, SciFin raise $44M for an agent "context layer" for enterprise revenue teams, and Empirik raise $21M for agents that predict enterprise infrastructure failures. (Sources: TechCrunch, TechStartups)

And Anthropic opened a research preview of its Model Hardware Standard (MHS), letting AI agents operate lab and manufacturing equipment directly — with Genentech, CMU, AWS and eight hardware vendors already integrating. It's the first serious push to extend agent standards from software into physical device control. (Source: Anthropic)

One legal cloud worth tracking: Sony Music Publishing and Warner Chappell sued Anthropic, alleging systematic scraping and torrenting of copyrighted works to train Claude. No damages figure has been specified yet, but it adds to the IP litigation risk sitting underneath the foundation models that enterprise agent platforms are built on. (Source: TechCrunch)

By the numbers

• 350%: quarter-over-quarter growth in agentic interactions across Cisco's 90,000-employee MyAgent rollout
• $805M: Okta's Q2 FY2027 revenue (+11% YoY), with guidance raised on AI agent identity demand
• 81%: B2B marketing leaders juggling 2+ disconnected AI tools weekly, per Optimizely's survey of 2,000+ leaders
• 27%: share of AI agent add-ons/skills AIR Security says it filters out as unsafe across its customer base
• 4 of 8: GitSpawn vulnerabilities across major AI coding agents still unpatched at disclosure

Read across

The throughline this week is that agent security stopped being a research topic and became a checkbox on procurement forms. Three security vendors (CrowdStrike, Proofpoint, Tenable) built agent-specific products in the same seven-day window, a fresh funding round (AIR Security) is being justified largely by the fact that a quarter of agent add-ons fail vetting, and OpenAI had to formally gate a model release for the first time on cybersecurity capability grounds. None of this is happening in a vacuum — it's catching up to real incidents (GitSpawn) and real enterprise spend (Okta's guidance raise). We wrote previously about how agentic AI differs from traditional RPA in what actually changes for enterprises (see the archive for that piece), and this week is a good illustration of one of the biggest differences: RPA never needed its own SOC vendor category. Agents apparently do.

If this was useful, forwarding it to one colleague who'd care is the best way to help it grow. Subscribe: https://buttondown.com/TheAutonomousEdge

Looking for more newsletters like this one? Discover more reads like this

Next issue: Monday, September 14, 2026.