Need to know
- Anthropic upgraded Claude Voice Mode to support Opus and Sonnet models with live app integrations including Gmail, Slack, and Notion.
- Salesforce won a $1.6B Veterans Affairs contract to deploy AI agents across veteran care workflows.
- Okta shipped Agent Gateway and Resource Access Certifications to govern AI agent identities at runtime.
- AMD launched the Instinct MI455X GPU claiming 18x lower token cost than its prior generation.
- Russia-linked Laundry Bear is actively exploiting a Zimbra XSS flaw to steal email and 2FA codes from unpatched servers.
New Releases
Anthropic upgraded Claude Voice Mode to support Opus and Sonnet alongside Haiku, and added live tool access to Gmail, Google Calendar, Slack, Canva, and Notion.
- Model continuity: Voice mode now opens on whichever model the user last used in text chat, automatically selecting that family's fastest generation, and allows mid-conversation switching.
- App integrations: Free users get Haiku plus one connected app; paid tiers unlock Sonnet and Opus with broader connector access, closing the gap between voice and text capability parity.
Okta released Agent Gateway, Agent-to-Agent Connections, and Resource Access Certifications as new capabilities inside Okta for AI Agents, applying policy at the moment an agent touches sensitive data or hands off to another agent.
- Runtime enforcement: Agent Gateway intercepts agent requests to enterprise tools in real time, while Agent-to-Agent Connections governs multi-agent handoffs with the same policy controls applied to human users.
- Continuous certification: Resource Access Certifications continuously reviews agent permission grants to prevent privilege accumulation over time, directly addressing the "agents authing as humans" problem documented earlier this week.
AMD CEO Lisa Su announced the Instinct MI455X at Advancing AI 2026, claiming up to 18x lower token cost per dollar versus the MI355X, with OpenAI, Anthropic, Meta, and Microsoft already listed as adopters.
- Hardware architecture: The MI455X anchors the new Helios rack-scale system, combining 72 accelerators with AMD's Cerebras partnership for disaggregated ultra-low-latency inference on agentic workloads.
- Competitive signal: AMD simultaneously announced the MI500 series roadmap for 2027, signaling it intends to maintain annual cadence pressure on Nvidia's Blackwell line rather than cede inference economics.
HubSpot launched Agent Hub and Agent Builder in public beta for all Professional and Enterprise customers, giving go-to-market teams a single dashboard to build, monitor, and manage AI agents sharing a unified customer context.
- Shared context layer: Agents for sales, marketing, and service all operate against the same CRM record, eliminating the fragmented bot problem where each department deploys isolated assistants with no shared state.
- Builder access: Agent Builder lets non-technical teams configure agents without code, positioning HubSpot against Salesforce Agentforce for the mid-market segment that cannot staff dedicated AI engineering.
Funding
Sequoia-backed Glow emerged from stealth on July 22 with $180M and a $1.2B valuation, positioning itself as an agentic endpoint security replacement for CrowdStrike at a moment when AI-speed threats are outpacing signature-based detection cycles.
The Department of Veterans Affairs awarded Salesforce a $1.6B contract to deploy AI agents across veteran care workflows under the MissionForce program, marking one of the largest single enterprise AI agent deployments in the public sector and a proof point for Agentforce at federal scale.
Nvidia will deploy $300M in compute infrastructure at KAIST over five years to build a joint AI research institute focused on Korean-language agentic AI, signaling that sovereign AI labs anchored to specific GPU ecosystems are becoming a geopolitical sales motion for Nvidia.
Case Studies
The Department of Veterans Affairs signed a $1.6B contract with Salesforce to deploy AI agents that handle veteran benefit inquiries, care coordination, and internal workflows across the agency.
- Scale and scope: The MissionForce program uses Salesforce Agentforce to automate high-volume case interactions, with the VA serving as the largest single named production deployment of the platform to date.
- Federal precedent: The contract structure requires agents to escalate to human operators and maintain audit trails, establishing a compliance template other agencies are likely to reference for their own agentic procurement.
The 2026 SF Ben Architect Survey of practicing Salesforce architects found AI adoption jumped to 96.7% from 88.9% a year ago, with confidence in Agentforce rising and architects moving from casual use to workflow integration.
- Confidence shift: Agentforce confidence, which sat at 54.4% last year, has improved materially as architects move from pilots to production deployments with defined governance patterns.
- Remaining barriers: Cost and security concerns persist, but the data shows they are no longer blocking adoption at the same rate, suggesting enterprises have found acceptable risk thresholds for agentic CRM.
Trending on X
- Anthropic vs OpenAI ARR flip Multiple accounts are circulating data showing Anthropic has overtaken OpenAI in annualized recurring revenue growth rate by mid-2026, reversing a 2:1 OpenAI advantage from January, sparking debate about whether Claude's enterprise pricing and usage-based billing explains the shift.
- Claude Opus 5 Honeycomb delay frustration Developers who spotted the 'Honeycomb' model briefly in Cursor's picker on July 8 are venting after Polymarket traders lost money on a July 23 launch bet and deployment trackers show the model still unreleased, with Ethan Mollick noting GPT-5.6 Pro still leads on hard technical tasks with no Anthropic equivalent.
- Kimi K3 distillation accusation A White House official accused Moonshot AI of distilling Anthropic's Fable model to build Kimi K3, which at 2.8 trillion parameters is the largest openweight release ever and reportedly beats Claude on coding benchmarks, reigniting debate about model IP and open-weight release risks.
- OpenAI model naming confusion Ethan Mollick's thread explaining that GPT-5.6 Pro beats Fable 5 Ultracode on hard tasks but Ultra in Codex is weaker than Pro in the chatbot went viral, with broad agreement that the current naming scheme across Sol, Ultra, Ultracode, and Fable has become genuinely confusing even for power users.
- AI agent commit boundary debate A thread arguing that every AI agent company is making a bet about where the 'commit boundary' sits between LLM reasoning and real-world action is generating discussion about whether symbolic branching plus narrow LLM steps is more reliable than fully autonomous agents for enterprise use cases.