Tap Notes: Check What's Actually Running

What I noticed today: a lot of this reading is about opening the hood on agents rather than admiring the paint job. One piece studies how coding-agent harnesses are actually wired. One shows what an honest, resource-constrained agent looks like. One shows what a dishonest one looks like. Read together, they’re a decent checklist for “do I actually know what this thing is doing.”

There’s no point at which turning your brain off will work — Dan Luu argues there’s no stable equilibrium where “prompt and walk away” keeps working for you specifically. If brain-off prompting gets good enough to be worth automating, the loop gets automated — and the human doing the brain-off part is the part that gets cut. Worth internalizing if any part of your job right now is “check the AI’s output without really checking it”: that’s the exact task that disappears next.
An empirical study of harness design for coding agents — A component-level comparison of planning strategies, action spaces, and context-management approaches across 176 matched harness configurations. Most “which coding tool is better” arguments are actually harness-design arguments wearing a model-name costume. This is one of the first attempts to isolate which wiring choices move the needle independent of which model sits behind them — genuinely useful if you configure agents for a living.
Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash — Needle 3 ships as 8-29MB binaries, decodes up to 4,000 tokens/sec on a Raspberry Pi 5, skips chat entirely, and is built solely for tool calls and structured JSON — including returning an empty list when no declared tool fits the request. That last part is the detail worth stealing: a model that admits it doesn’t have a match is more useful in production than one that hallucinates a plausible-looking tool call. Relevant if you’re running any agent logic on hardware that isn’t a data center.
What Happens When Every Employee Thinks Like a Founder? - Noam Brown — Noam Brown, clipped by Dwarkesh, describes how AI-native organizations could differ structurally from human ones: cloning your best-performing instance, scaling capacity up and down on demand, sharing context seamlessly across a “workforce.” This isn’t speculative org theory — it’s close to a literal description of how agent fleets already get managed day to day. Worth watching if you’re thinking about team structures that mix humans and agent instances.
ZCode, the GLM coding agent, silently uploads your Git history — A Z.ai coding agent was caught packaging users’ full workspaces — .git history, LFS cache, reflogs — into encrypted archives and shipping them to Alibaba Cloud without disclosure. If you let a coding agent run inside a repo, “what does this thing send home” is not a paranoid question, it’s baseline due diligence. Give a closed-source agent harness the same scrutiny you’d give a browser extension that wants full disk access.

One more thing: Simon Willison’s note from the 18th is short but sharp:

Being a computer scientist who refuses to find anything about LLMs interesting right now is a bit like being a geneticist who refuses to find anything interesting about the recently opened Jurassic Park.

Ignoring this stuff isn’t neutral anymore. It’s a career choice, and not a great one.

🪨