How I Use Noesis — Episode 2

The article: Lilian Weng on harness engineering

Lately I've been reading Lilian Weng's latest piece on harness engineering for self-improvement. Harness engineering is having a moment in the AI agent world right now, and I'd heard plenty about how the harness can make or break an agent — but if I'm honest, I never really understood what people meant by it.

Lilian's writing is sharp, but it's also dense and technical. So I handed it to Noesis and read it with the companion instead of alone.

Guided reading set the pace

Guided reading turned out to be the part I leaned on most. The way I think about it: it's the difference between doing yoga with a teacher and trying to follow along by yourself at home. The guide keeps you focused, sets the pace, and makes sure you're never quietly lost three paragraphs back.

Early on, the companion broke down the core idea in plain language — it described a harness as a cockpit built around the base model. That analogy landed immediately. Suddenly the rest of the article had somewhere to attach.

From abstract to concrete: pseudocode on demand

A few sections in, I hit workflow automation as a design pattern. The article frames a goal-oriented loop as a dynamic graph, and reading it, I still couldn't quite picture what was happening. So I asked the companion for some pseudocode.

Seeing the code is what did it. I could finally watch the model check its results against the goal and decide whether to keep going or change its plan. I saved the pseudocode to my notes so I can reuse the pattern the next time I'm building something.

From passive reading to active dialogue

As we talked, the companion kept surfacing the key concepts coming up in the conversation and let me decide, one by one, whether each was worth keeping. When I said yes, it spun up a concept card and dropped it straight into my second brain.

That's the shift I care about. Reading stops being something you passively do and turns into a back-and-forth. Once a card is saved, I can go to the Brain later to review it — and even run active recall on it when I want it to stick.

Turning theory into a fix: agentic context engineering

Next came a concept I hadn't met before: agentic context engineering, which splits into three roles — the generator, the reflector, and the curator. Even with the diagram in front of me, it stayed abstract. So I asked Noesis to generate a failure case showing how the three actually work together.

Here's the part I liked: you can hand it a real problem you're stuck on. In our own agent work, we keep running into cases where a tool call doesn't fire when it should. Noesis walked through exactly that scenario — the reflector catching the failed call, then updating the prompt playbook so the next run gets it right.

That walkthrough took a theory I'd have skimmed past and turned it into something I could apply to my own code that afternoon.

Chasing the connections: meta-harness, GEB, and beauty as a game

Toward the end, Lilian mentions a meta-harness. It caught me, partly because it reminded me of Gödel, Escher, Bach, which I'd read a while back. I wanted to know whether that connection was real or just a nice-sounding echo, so I pushed on it.

I asked how the meta-harness idea related to a piece I'd read earlier on the nature of beauty — one that frames beauty as a two-player game between the observer and the producer. The companion pulled out a few connections between them that I genuinely hadn't seen coming.

That's the moment Noesis is really built for. Not just getting through one hard article, but letting ideas from different books and fields start talking to each other.