AI Daily Recap

The day's center of gravity was the scaffolding around AI agents rather than the models themselves. Anthropic published a guide to building with Claude managed agents, a fresh crop of open-source projects tackled the unglamorous parts of running agents in production — long-running goals (Interbase), isolated per-branch databases (SafeAgentDB), and cost visibility (AgentMeter) — and the tooling story stretched from the IDE, where Xcode 27 folded Gemini into its agentic coding toolset, down to the silicon, where AWS detailed agents that hand-tune Trainium kernels.

Underneath the agent layer, infrastructure had a busy day. Microsoft open-sourced pg_durable to run durable workflows natively inside PostgreSQL, Azure API Management shipped a unified model API plus content-safety policies that now cover MCP tool calls, and a new Python library, llmbuffer, went after prompt-cache hit rates for big cost savings — a reminder that as agents proliferate, the economics and plumbing get most of the attention.

Models weren't entirely quiet: Simon Willison flagged DiffusionGemma, Google's return to diffusion-based language models after last year's Gemini Diffusion preview, and Latent Space covered the aftermath of Anthropic's Fable 5 launch and its controversial usage terms. Jeremy Howard, meanwhile, offered the day's sharpest provocation on how labs might voluntarily slow recursive AI self-improvement.