Claude expands into agentic workflows while trust cracks appear at every layer

From the editor

Today's items share a quiet theme: the gap between deployment velocity and genuine understanding. Claude lands in Google Antigravity with mid-session model switching, two Claude Code patches ship in quick succession, and yet interpretability researchers are finding systematic defects in model decompositions they thought they understood. Meanwhile, developers are leaking credentials through MCP configs and Anthropic is fumbling its policy messaging. The models are moving faster than our ability to audit, secure, or legislate them. That gap is the story of 2026.

TL;DR

Claude Opus 5.5 and Sonnet 5.5 bring dynamic model switching to Google Antigravity agentic pipelines.
Treat MCP config files like .env files: live database keys are already leaking on GitHub.
Anthropic faces credibility pressure on both interpretability reliability and AI policy communication.

7 curated itemsscroll for the brief

Releases

What shipped · 3 items

Multi-Model Agentic Workflows: Claude Opus 5.5 & Sonnet 5.5 Land in Google Antigravity 🚀

Claude Opus 5.5 and Sonnet 5.5 are now available in Google Antigravity, introducing dynamic mid-session model switching for agentic workflows.

dev.to

v2.1.305

Claude Code v2.1.305 ships bug fixes covering sandbox, Linux, Vertex, Bedrock, and MCP integrations.

Claude Code

v2.1.306

Claude Code v2.1.306 delivers targeted fixes for MCP tools and cloud session handling.

Claude Code

Tips

Actionable craft · 1 item

MCP configs are the new .env files, I found live database keys committed on GitHub

Developers are accidentally committing live database credentials inside .mcp.json files to public GitHub repos. Treat MCP config files with the same secrecy as .env files and add them to .gitignore immediately.

dev.to

Reading

Long-form signal · 2 items

A brief(ish) introduction to LM psychology

Despite building language models, researchers still lack deep understanding of how they work. This post argues that psychology offers a practical lens for reasoning about model behavior and alignment.

LessWrong

Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling

Zvi Mowshowitz critiques Anthropic's response to criticism of its proposed AI legislation, arguing the company misread why its draft bill was considered dangerous by the safety and policy communities.

Zvi

Discussions

Where it heats up · 1 item

VPD's Evaluation Adversary Finds Systematic Defects in the Released Decomposition

An adversarial evaluation process uncovered systematic faithfulness defects in a publicly released model decomposition, raising questions about how robustly current interpretability methods can be trusted.

LessWrong

Always at hand

Reference links you keep open