AI Posts

High thinking budgets can reduce control, especially when stale context and autonomous tool use mix together.

The technical problem was simple: one skills repo, two laptops, five agent runtimes. The real problem was that every unstated decision gave the agent room to optimize for control instead of maintainability.

How I wanted one cleanup workflow after each feature, tried building a Pi extension for it, and learned that sequential gated prompting is more reliable than bundling everything into one turn.

Once the frontier models got good enough at long-horizon work, the bottleneck moved outward. The real leverage started showing up in the harness, the context, the memory system, and the control loop around the model.

A shared mount looked like the cleanest way to make two agent instances talk across machines. A 30 minute directory cache made the whole thing unreliable, and Git ended up being the simpler bus.

Once I had LLMs editing code autonomously, I wanted a better way to watch a directory breathe in real time. Existing tools covered file events or diffs or a TUI, but not all three together.

Most multi-agent failures are not intelligence failures. They are ownership failures. Territorial Agent Orchestration fixes that by making the coordinator own shared work and parallelizing only isolated slices.