SoN 2.37: Dumb Little Robots, All the Way Down

VOL 2, ISSUE 37 · 5 OCTOBER 2026

It's been two weeks since the last issue. Life has got in the way, and family has taken precedence.

This issue leads with two pieces about the same idea: break the work down, and give each piece what it needs. Dumb Little Robots is about an agent that got confused because it was given every job at once. The difference between skills and context is about separating the instructions for a job from the material that goes into it.

Dumb Little Robots, All the Way Down

A team I spoke to recently built an analyst agent and gave it the whole job: bring the data in, analyse it, report on it. Its operator barraged it with questions. Its brain was full. The context window filled, the tokens were used up, and it lost track of what it was for.

Anthropic's guide to context engineering describes a related effect: "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases." Chroma tested 18 models and found performance "grows increasingly unreliable as input length grows."

I told the team to think of agents as dumb little robots all the way down. One agent pulls data from a domain, one normalises it, one analyses it, one formats the data, one formats the PowerPoint. An agent might take two of those jobs, but you don't give it everything in one go.

I covered decomposition in The Art of Breaking Things Down and it needs repeating, because enterprise teams are still being sold AI as the number one problem solver you can throw everything at.

Five small agents means four handoffs, and someone has to check them. Each agent needs a person who owns it, understands it and takes responsibility for how it performs. Any system a team builds should have that accountability anyway.

These things are only as good as we program them to be. An agent that gets confused may simply have been given too many jobs.

Read it on the site →

The difference between skills and context

On 2 October I posted Stupid Gemini Tricks, a Gemini Gem I built for a U of Digital AI accelerator training. It makes a one-panel cartoon in the style of the New Yorker or The Far Side, and it took about two hours and a lot of iteration to get the prompt and context right.

The skill is the doing: the instructions for building the thing. The Gem's instructions say "Act as a veteran magazine gag cartoonist and creative editor," then tell it to pitch two or three concepts or punch up a gag the user already has, confirm the style, and then generate the image with the caption beneath the panel. Anthropic's documentation for Agent Skills says that "unlike prompts (conversation-level instructions for one-off tasks), Skills load on demand, so you don't have to repeat the same guidance across conversations."

The context is the contents: the things that go into a particular job. The Gem carries a style reference guide, with the New Yorker's "dip-pen and black ink with soft watercolor/wash shading" and the Far Side's "1980s retro muted print tones." The user brings the topic or the gag.

The context is the ingredients on the counter, the skill is the recipe card, and the model is the cook.

Anthropic defines skills as "reusable, filesystem-based resources" that hold "workflows, context, and best practices", so the style guide can travel with the skill. When Claude loads a skill, its instructions enter the context window too.

Read it on the site →