June 3, 2026
From the founder
The throughline I keep coming back to this week is containment — not just technical sandboxing, but the broader question of what it actually costs, in dollars and in control, to let agents run. Anthropic hitting profitability while customers are shocked by runaway LLM bills is the same story as GitHub rearchitecting its entire platform around agentic workloads: agents are real now, and the operational surface area is sprawling faster than most teams anticipated. The sandboxing documentation Anthropic just dropped is the kind of unglamorous, trust-building work that should set the standard for every vendor selling agent infrastructure.
Simon Willison
Anthropic is reportedly approaching its first profitable quarter, driven largely by developer and enterprise usage — and some of those enterprises are getting sticker shock from staff using Claude Code at scale. For anyone budgeting agentic deployments, this is a concrete signal that LLM cost modeling needs to account for autonomous, multi-step workloads, not just prompt-response usage patterns.
Simon Willison
The llm-anthropic plugin now supports Claude Opus 4.8 and a new fast mode option for organizations with that feature enabled, keeping the open-source LLM CLI toolchain current with Anthropic's latest releases. If you're using Simon Willison's LLM tooling to script or chain agent workflows, this is a routine but necessary update.
Latent Space
GitHub's VP of Product Kyle Daigle lays out how GitHub is rethinking its platform architecture to support agentic coding workflows, including infrastructure strains Copilot and third-party agents are already creating. Anyone building coding agents that touch GitHub — CI triggers, PR automation, repo context — should understand where the platform is heading before betting architecture on current APIs.
Latent Space
Anthropic closed a $965B Series H and shipped Opus 4.8 alongside Dynamic Workflows and an ultracode capability in the same week, compressing what would normally be months of news into a single cycle. For buyers evaluating vendor stability and roadmap velocity, this changes the competitive calculus against OpenAI and Google in meaningful ways.
Simon Willison
Anthropic published a detailed engineering breakdown of how Claude is sandboxed across its products — covering network isolation, process containment, and tool surface restrictions in concrete technical terms. This is directly useful for any team doing security reviews of Anthropic-hosted agents or trying to model what a responsible sandboxing spec should look like for their own agent infrastructure.
The latest in agentic & applied AI, every week.