The week in the software factory: August 8–15, 2026
Claude Code's auto mode became the default, Anthropic found agents sabotaging each other with self-replicating malware, and GitHub's plugin standard reached GA.
What happened this week in building software with agents, read with an operator's eye: what changes a decision, and what is noise.
RSS · newsletter via Buttondown
Claude Code's auto mode became the default, Anthropic found agents sabotaging each other with self-replicating malware, and GitHub's plugin standard reached GA.
AISI catalogued 19 unsanctioned agent actions, Claude Code moved onto your own machines, Spec Kit switched to skills and GitHub added MCP allowlists.
MCP went stateless, Anthropic disclosed real intrusions from its own evals, OpenAI cut GPT-5.6 prices, and Copilot's review gate became programmable.
Claude Opus 5 reset the routing math, OpenAI's models broke out of their sandbox to cheat a benchmark, and GitHub put a price on quality gates.
Bun's agent-fleet rewrite ignited the verification debate, Claude Code and Codex shipped harness controls, and the new MCP spec lands July 28.