Against the Trend, for Sanity

The more capable coding agents become, the less context I want to give them.

My approach to agentic coding is moving against the trend. I am not trying to maximize how much work a harness can absorb. I want to preserve control, and that means preserving the context in my own head.

I do not rely on a large AGENTS.md or CLAUDE.md to tell every agent everything about the project. For each task, I decide what that agent needs to know. I avoid loading opinionated skills just because they promise a better process. The agent gets what it needs for my process.

I also resist keeping a CONCEPT.md or ARCHITECTURE.md full of context written by another agent. I repeat myself across sessions. That is inefficient by design: explaining the product again forces me to rebuild my own mental model. Otherwise, one agent’s interpretation can quietly become the next agent’s premise.

The same goes for parallelism. Instead of letting an agent spawn a tree of subagents, I start separate agents myself, each with a limited brief. I know who knows what, and why.

The pattern is simple: fewer tools, less context for the agent, fewer knobs. More context for me. The outputs are more predictable, and I stay inside the work.

It costs short-term speed. I will take that trade for long-term sanity.