Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic

Anthropic's Applied AI team presents how agents — AI systems that independently solve tasks — should be built to handle long-running production jobs.

They show that outdated solutions become overhead as models become smarter: Sonnet 4.5 was cautious with memory and exited early, so they built in a fix — but when Opus 4.5 was deployed without this behavior, the fix became merely slow redundant code. The key insight is separating the agent's brain (reasoning loop) from its hands (tool execution), making systems 60–90 percent faster and making errors recoverable instead of fatal. A log of everything that happens during a session serves three purposes: seeing what the agent does, retrieving information Claude forgot, and something they call "dreaming" — a nightly job that improves the agent's memory for the next day.

More to read

Lumaril, which develops AI software for optimizing industrial production by connecting to existing control systems, raised a $66M Series A at a $400M valuation

Techmeme · 3 h ago

DeepSeek open sources Super-Agent as a research prototype, a proactive coding agent it says can monitor group chats for coding issues and suggest solutions

Techmeme · 4 h ago

Insurance software maker Guidewire agrees to acquire Hyperexponential, which makes AI-powered insurance underwriting and pricing software, for $1.7B in cash

Techmeme · 4 h ago

One word for AI agents in 2027

Nate Herk · 8 h ago