what a skee is

A very small agent with a very short list of powers.

A skee is not a chatbot with a wallet. It has no conversation, no memory of you, and no goals beyond the one errand you gave it. That is not a limitation we're apologising for. It's the design.

Why "tiny" is the whole point.

Most agent products start from a large model and try to bolt on safety. The model can do anything, so the product spends its life trying to stop it from doing the wrong thing.

skee. starts from the other end. A skee can do one errand, chosen from a library of small, audited programs. It cannot improvise. It cannot decide that today it will try something new. If you want a different behaviour, you spawn a different skee.

The result is an agent you can leave alone for a month, which is the only kind of agent that is actually useful.

The three parts.

job

The errand and its parameters. "DCA $25 into TSLAx." "Keep 60/40." "Sweep balances under $1." One sentence, fully specified, no ambiguity for a model to fill in.

leash

Hard limits enforced by the runtime at signing time: spend cap, allowed assets, allowed hours, slippage band, kill switch. The skee cannot exceed them even if its errand code is wrong.

nap

When it wakes. A skee is asleep almost always. It does not watch the market, it does not "monitor". It wakes on a schedule or a trigger, runs once, and sleeps again.

What a skee is not.

a typical "AI agent"a skee
decides what to doa model, at runtimeyou, once, at spawn
can be promptedyesno — there is no prompt
can exceed its budgetif the model is convincednever — runtime refuses to sign
runscontinuouslya few seconds, then sleeps
explains itselfin paragraphs, after the factone plain-English receipt per run
number of jobsmany, at onceexactly one

Where the intelligence actually lives.

Errands are written by people and reviewed before they enter the library. Some errands use a model to classify — for example, "does this new listing match the filter 'US equity, large cap'?" — but the model's answer is only ever an input to a rule, never a permission to act.

The runtime holds the keys. The runtime checks the leash. The runtime signs or refuses. Nothing a model outputs can change that path.

Rule of thumb: if a skee ever surprises you, that is a bug report, not a feature.