AI Agent Infrastructure | AI Field Notes #35

AnalysisAnthropic reported on May 20 that it is on track to record $10.9 billion in Q2 2026 revenue, more than double the $4.8 billion from Q1, and projects $559 million in operating profit. That would be the company's first operating profit, arriving two years ahead of internal projections. The accounting has significant asterisks: the figure excludes stock-based compensation, the company is not subject to public disclosure standards, and the two profitable months coincide with a ramp-up discount on a new compute deal with SpaceX. That deal is worth $1.25 billion per month once the full rate starts in July, for access to SpaceX's Colossus supercomputer infrastructure through 2029. Anthropic's own guidance acknowledges the company may not remain profitable through the rest of the year once full compute spending resumes.

AnalysisIn late May, users spotted a toggle for "claude-mythos-1-preview" in the public Claude Code interface before Anthropic removed it. Mythos was announced April 7 as a restricted frontier model built for computer security. During internal testing it found 10,000 high- or critical-severity vulnerabilities in its first month and demonstrated the ability to generate functional cyberattacks described as professionally capable. Public release has been delayed while guardrails are developed. Anthropic simultaneously announced "Glasswing," a collaboration with more than 50 organizations to surface AI-driven vulnerabilities before they reach infrastructure. The Claude Code toggle suggests a staged rollout is being prepared, with the company treating the transition as a controlled beta rather than a general release.

AnalysisAnthropic's developer conference in early May brought a series of Claude Code updates that shift the tool from an interactive coding assistant toward a semi-autonomous development environment. Worktrees let Claude create isolated git branches and work in them independently without touching the user's active branch. Routines execute saved prompts on scheduled intervals or via webhooks, running without a human session open. Remote control enables a session started on one machine to continue on a mobile device. Auto mode runs classifiers to screen for destructive actions before executing them. Managed Agents, now in public beta, adds sandboxed code execution, checkpointing, and credential scoping as primitives for building longer-running pipelines.

AnalysisGoogle CEO Thomas Kurian confirmed at I/O 2026 that Gemini technology forms the foundation of Apple Foundation Models, the engine behind a redesigned Siri set to launch later in 2026. The deal, first reported in January and valued at approximately $1 billion per year, came after Apple tested OpenAI and Anthropic alternatives before settling on Google's technology. The new Siri will run primarily on Apple devices and Private Cloud Compute, with Gemini providing the model base. The arrangement positions Google as the AI backbone for its own products and for Apple's roughly 2.2 billion active devices, making it the largest single consumer AI distribution arrangement in the industry.

AnalysisAnthropic disclosed that a packaging error in April caused internal TypeScript files from Claude Code to be bundled into a public distribution before the issue was caught. The leaked material included references to internal features, permission systems, orchestration workflows, and sandboxing mechanisms. Security researchers noted on May 25 that the exposed internals give potential attackers a clearer map of decision boundaries and bypass surfaces inside the tool. Anthropic characterized the exposure as unintentional. No evidence of active exploitation has been reported, though the incident arrives as Claude Code is being extended with Managed Agents infrastructure, MCP tunnels, and multi-device remote sessions.

AnalysisCisco announced it would cut approximately 4,000 jobs in Q4 2026, its second significant layoff round this year, bringing the 2026 total to roughly 9,000 employees. The stated rationale was a focus on AI, silicon, optics, and cybersecurity. In the same period, Cisco reported $5.3 billion in AI-related product orders, fueled by data center networking demand from companies expanding AI infrastructure. The combination captures the pattern repeating across large tech companies: the AI buildout generates hardware and connectivity revenue while reducing the need for the workforce that managed the previous generation of systems.