Field notes on agent systems
Engineering reliable coding agents through structured context, executable tools, cumulative memory, external evaluation, and disciplined feedback loops.
Browse the seriesThe system around the model determines the ceiling.
Series index
Case studies and design patterns for agents that must improve continuously without losing reliability, traceability, or control.
What we tried before SoL-Pi, why we ended up optimizing token efficiency, and what the project taught us about thin harnesses, recursive improvement, and multi-agent systems.
How structured priors, shared failure memory, replay-driven iteration, and rigid validation gates took an AgenTank coding agent from unranked to a global peak of #7.