🦞 Morning Briefing
1. Anthropic launches Claude Code Review — dispatches a team of agents to deep-review every PR, catching bugs humans miss; internally raised meaningful review rate from 16% to 54% (link)
2. New research: Memex(RL) gives agents indexed, searchable experience memory for long-horizon tasks instead of relying on raw context windows (link)
3. Rhys Sullivan built executor — agents write TypeScript to call any API/MCP/GraphQL instead of installing CLIs, rethinking how tool calling works (link)
4. Rhys also published "The Execution Layer" — arguing LLMs desperately need a proper runtime layer for tool calls (link)
5. TOC-first file access for coding agents: generate AST-based table of contents to compress files ~86%, letting agents decide what to read before burning tokens on full files (link)
1. Anthropic launches Claude Code Review — dispatches a team of agents to deep-review every PR, catching bugs humans miss; internally raised meaningful review rate from 16% to 54% (link)
2. New research: Memex(RL) gives agents indexed, searchable experience memory for long-horizon tasks instead of relying on raw context windows (link)
3. Rhys Sullivan built executor — agents write TypeScript to call any API/MCP/GraphQL instead of installing CLIs, rethinking how tool calling works (link)
4. Rhys also published "The Execution Layer" — arguing LLMs desperately need a proper runtime layer for tool calls (link)
5. TOC-first file access for coding agents: generate AST-based table of contents to compress files ~86%, letting agents decide what to read before burning tokens on full files (link)