Field Notes
Tourist loads a repository as a city, one building per file, and runs coding agents against it. Scoring, memory, and a reinforcement-learning policy sit on the same run. A real pull request is stage 8.
From a 20-minute delivery estimate to LLM latency: explore how a line learns, why we square its mistakes, and when a good-looking prediction stops being useful.
GPT-6 Astra is not just a stronger reasoning model. Its computer-use gains, critical cyber capability, and monitorability limits change how we should design and supervise AI agents.
A practical design for bounded agent loops: finish when the work is good enough, compact before context degrades, and stop safely when progress stalls.
A practical architecture for agents that turn experience into safer, measurable improvements through memory, sandboxed experimentation, evaluation, and replay.