From Code Steering to CrowdStrike: Agents Enter the Defense Layer

Research

Google DeepMind's Code Steering with MCTS for Precise Agent Control

This tackles the core problem of getting agents to do exactly what you want in complex software. Instead of firing in the dark, it uses Monte Carlo Tree Search to systematically explore action sequences. For builders, this is a direct tactic for improving reliability in real-world automation without hitting a wall.

Tools

Claude-Code gets a "Context Map": Localizing the Agent Mindset

We've spent months building universal protocols, now the focus is shifting to local reliability. Claude Code is forcing the agent to stay grounded in the local project context using context maps. This prevents the agent from hallucinating global states when you're actually trying to debug a single repository.

News

MCP Auth Spec: The Missing Piece for Secure Cross-Service Agents

The most important part of the MCP spec debate right now is this PR regarding authentication. Agents need to act on our behalf across different Web APIs, and this proposal standardizes how they securely authenticate to do so. Without this, integration will remain a fragmented security nightmare.