Need to know
- Nvidia's Groq 3 LPX enters full production at Samsung, delivering 3,431 tokens/second on the Vera Rubin platform.
- Okta Agent SSO reaches general availability, giving AI agents the same identity controls as human employees.
- Anthropic ships enterprise-managed MCP connectors, turning Claude's tool access into a policy-enforced identity layer.
- Meta plans to launch consumer AI agent platform Hatch as early as September, with model Watermelon following in October.
- Clay adds Account Agents to workflows, letting GTM teams embed account research directly into automated plays.
New Releases
Okta made Agent SSO generally available, extending its Cross App Access protocol so AI agents are managed as first-class identities under the same policies as human users.
- Static API keys replaced: Agent SSO eliminates high-risk static credentials and repeated consent prompts by issuing agents short-lived tokens governed by centralized Okta policy.
- Cross App Access standard: The XAA protocol is already embedded in Okta's platform serving 20,000-plus customers, meaning Agent SSO inherits existing policy scopes without new infrastructure.
Anthropic released enterprise-managed controls for Claude's MCP connectors, letting IT and security teams govern which tools agents can access rather than leaving that to individual users.
- Admin-layer authorization: Organizations can now centrally approve, restrict, or revoke connector access across Claude deployments, treating MCP authorization as a security policy rather than a user preference.
- Identity-first framing: The release positions MCP management alongside Okta-style access controls, signaling that Anthropic views connector governance as an enterprise security surface, not a product feature.
Nvidia confirmed its Groq 3 LPX inference accelerator is now in full production, manufactured by Samsung, with first deployments heading to neocloud provider Nebius later this year.
- 3,431 tokens per second: Benchmarked by Artificial Analysis on Gemma 4 31B with a 100K context window, Groq 3 LPX runs four times faster than the nearest competing inference chip.
- Vera Rubin platform integration: The LPX rack pairs with Nvidia's Vera CPU and Rubin GPU, completing an inference stack purpose-built for agentic workloads that consume up to 15x more tokens than chat.
Clay shipped Account Agents inside its workflow builder, letting revenue teams embed autonomous account research directly into GTM plays rather than treating it as a separate manual step.
- Action inside the workflow: Account Agents can now execute research and take downstream actions within an automated sequence, collapsing the gap between data enrichment and outreach.
- GTM signal integration: The same day, Clay confirmed an Intentsify integration bringing account-level and persona-level intent signals into the platform.
Funding
Palo Alto Networks is reportedly in acquisition talks with one-year-old endpoint protection startup Koi at a $400M valuation, continuing an aggressive inorganic AI-security buildout that already includes CyberArk and Protect AI.
Mistral and Saudi Arabia's HUMAIN announced a strategic collaboration covering AI infrastructure, model development, and localized AI deployment across the region, signaling European open-source AI is now competing directly for sovereign AI contracts against US hyperscalers.
Case Studies
Anthropic shipped a frontend rendering optimization to Claude's web and desktop apps on August 24 that reduces stall events by 9x on slower laptops during long-form responses.
- No model change required: The improvement is a client-side rendering fix, not a model or API update, meaning all existing Claude deployments benefit without any integration work.
- Target use case: The fix directly addresses long-context agentic outputs, where token streams are longest and stutter has been most disruptive for enterprise users running extended workflows.
Trending on X
- GPT-5.6 Sol price cut debated Developers are parsing whether the 20% input and 33% output price cut through November 21 is a strategic move to block Anthropic or a signal that Sol's economics weren't landing at launch pricing.
- Meta Hatch agent launch imminent The AI community is tracking Meta's September target for Hatch, a consumer agent that interacts with DoorDash, Etsy, and Outlook, as the first real test of whether Meta can monetize its AI stack beyond the free tier.
- Okta Agent SSO and agentic identity control Security practitioners are debating whether short-lived agent tokens via XAA actually solve prompt injection and privilege escalation risks, or just add an identity wrapper to an unsolved problem.
- Anthropic withholding Mythos 2 model Reports that Mythos 2 is done training but not being released externally are generating discussion about whether capability suppression becomes standard practice as models cross new risk thresholds.
- Nvidia Vera Rubin 30x DeepSeek throughput benchmark Builders are scrutinizing Nvidia's first published Vera Rubin NVL72 benchmark showing 30x DeepSeek throughput, asking whether the gains hold on real agentic workloads or only on synthetic coding tasks.