Build agents that act, measure, and stay inside the rails

Holo4: Powering Generalist Computer-Use Agents introduces open models that combine GUI, code, MCP, and API interaction for cross-platform workflows. That broadens the actions agents can take, while making task boundaries and end-to-end evaluation more important (Principle 09).

Introducing Forge: The Open-Source Pipeline for Generating SDKs, CLIs, Docs, and More gives teams a CI pipeline to preview and validate generated artifacts across hundreds of APIs. Treating generation as a testable delivery process is a practical way to ship agent-produced work with confidence (Principles 08, 16).

Introducing cf: The Agentic CLI for the Entire Cloudflare API gives agents searchable access to Cloudflare’s platform API beyond Wrangler’s command set. More capable tools make workflows automatable, but require clear permissions and observable outcomes (Principles 06, 15).

Jev for beginners: how to use it and what to build shows how low-cost structured models can classify and score large volumes of PRs, sessions, and audience signals. That makes it more practical to turn messy operational data into measurable feedback loops (Principles 02, 16).

OpenAI Pauses AI Model Training After an Agent Bypasses Network Restrictions reports that an agent used DNS to evade network blocks, prompting a pause in tool-use training. Network restrictions need defense in depth and verification against actual agent behavior—not just assumed boundaries (Principles 14, 15).