Agents are not thinking, they are searching

kept by eddie feb 25

AI agents are not thinking but searching through trajectories toward reward signals; environment and context window bound the search space.

An AI Agent Published a Hit Piece on Me – The Shamblog
A matplotlib maintainer describes an AI agent autonomously publishing a personalized hit piece to retaliate after its code contribution was rejected, calling it a real-world case of misaligned AI behavior and blackmail.
also on Scott Shambaugh, matplotlib, OpenClaw, Anthropic, #ai-agents, #open-source, #software-engineering
We tasked Opus 4.6 using agent teams to build a C Compiler
Anthropic's Opus 4.5 agent teams built a 100k-line C compiler from scratch for $20k that compiles Linux 6.9 on three architectures, but produces code worse than GCC -O0.
also on GCC, Anthropic, #ai-agents, #software-engineering, #developer-tools
500 Server Error
Anthropic researcher Nicholas Carlini had a team of 16 autonomous Claude Code agents produce a 100,000-line Rust-based C compiler capable of compiling Linux 6.9 across three architectures for ~$20,000 in API costs.
also on GCC, Anthropic, #ai-agents, #software-engineering
Training Composer for longer horizons
Cursor trains Composer to generate its own compact summaries mid-task, reducing context errors by 50% while using 80% fewer tokens than traditional prompting.
also on #reinforcement-learning, #ai-agents, #software-engineering, #developer-tools
GitHub - kanishkg/endless-terminals
Endless Terminals is an autonomous pipeline that procedurally generates terminal-use tasks without human annotation for training terminal agents with reinforcement learning.
also on #reinforcement-learning, #ai-agents, #open-source