Own your agent and the context

Context bloat is getting ridiculous these days, and agent harness matters a lot for these clankers.

By “context bloat”, I mean something pretty concrete: excess or low-quality context can raise cost, hurt performance, and give the agent more chances to follow the wrong thing.

Harness effects are real. That is why I appreciate open source agents more than closed source ones. You need to know what the fuck is going on in there.

From what I can synthesize, most of the current research on context “engineering”(?) points toward less is more. Or, more accurately: less always-loaded, irrelevant, or duplicated context; more targeted, routed, and owned context when the task actually needs it.

Write your own context because it helps. Unchecked synthetic context can hurt and raise costs. Agents are too obedient, which sounds good until you realize they will also obey stale instructions, noisy skills, and bad project lore.

Imagine you are working on your backend and ALL the React skills and Docker context are loaded. Wasteful af.

That is why I tend to avoid /init, though that does not mean you should NOT use it as scaffolding for your AGENTS.md.

Most public skills are bloated noise. Cut the non-actionable body material and quality can rise. The point is not to make every skill tiny. The point is to remove whatever does not help the agent act.

I have noticed that most of these skills lack quality, and I am usually better off writing my own. I either fork, rewrite, or revise the content.

Descriptions need enough signal to route, but not so much that they get truncated or ignored by the harness. Short, specific, and discriminative beats long and vague.

Long reasoning chains can make agents forget what you told them. More “thinking” is not automatically better if the model drifts away from the original constraints.

So, I can say what Mario, the creator of Pi, is right, or at least I agree.

You must own your agent and the context.

I think this is much more useful than just prompting harder. Though it still depends(?) because some models are optimized for some harness, like Codex with recent GPT models. Still a “trust me bro” benchmark.

Again, we are still in the fuck around and find out (FAAFO) era of agent harness. One feature that makes Pi nice is the ability to allow the model to self-evolve and edit its own harness.

I tend to think that:

Windows is to Claude Code
Ubuntu is to OpenCode
Arch is to Pi

Regardless, do you know what the most critical thing is, above all this noise? Review your code, edit it yourself, and make sure you actually know what the fuck is going on.

Never be sycophantic with the tool, because the tool is already built to agree with you. If both of you are nodding along, nobody is driving.

Destination: Nabunturan
Agent: “Direction -> Panacan -> Bukidnon -> Cotabato”
You: “LGTM.”

And that is how you end up confidently going everywhere except where you were supposed to go.

I am still trying to learn how to leverage these tools without compromise (because skill atrophy is real), and learning them while reading the research helps me find what works for me. Most of the time, I read what the best engineers do rather than fucking follow the sloppers.

Most of the pioneers are already adopting this stuff, and they often document what they have learned along the way. And that is a gold mine! ^{[1] [2]}