What is an agent template?
An agent template is a folder that tells an AI agent what job to do and how to do it. Someone wrote it while doing that job for real, and published it so you can start from something instead of nothing. Here is what arrives, what you may change, and what we check before it reaches you.
The short answer
An agent template is a folder of instructions and skills that an AI agent reads before it starts work. The instructions say what the job is. The skills are the steps it runs to do that job. Download it, unpack it, and your agent has somewhere to start instead of an empty page.
You will also see the download called a pack. Same thing: the page you browse here is an agent template, and the file you take away is the pack.
- An agent that reads your supplier invoices and flags the three lines that do not match a purchase order.
- An agent that reads a support inbox and drafts replies in your voice, for you to approve before they go anywhere.
- An agent that opens a competitor's pricing page every Monday and tells you what moved.
An agent harness — the program that reads the folder and runs the agent. Hermes, Claude Code, Codex CLI, DeepSeek Harness, Cursor and OpenClaw are six of them, and the README says which one its author used. About ten minutes, and downloading needs no account.
What is inside one
A pack is small, and all of it opens in a text editor. Six things, usually:
| What you will find | What it is for |
|---|---|
| SOUL.md | who the agent is: its goal, its style, and what it is allowed to do |
| skills/ | one folder per skill, each holding the steps the agent runs |
| config/config.template.yaml | the model it talks to, and where its memory lives |
| README.md | the job, who wrote it, and how to start it |
| LICENSE | what you may do with it — MIT, most often (how to read a licence) |
| the version note | what changed in this build, and when |
SOUL.md is the one to read first. It is short, often a paragraph, and it is where a folder turns into a colleague:
Agent goal — Find lines that do not match a purchase order, and say why. Style — Terse. A table first, the explanation under it. No preamble. Authority — Read-only on the finance folder. Raise questions; never edit an invoice.
That last line is the guardrail, and it is the line worth reading twice in anything you download. The parts of an agent walks through everything else a template can carry: goals, guardrails, memory, self-improvement and reporting.
Who made it, and what we check
Every agent template names the person who uploaded it. They wrote the pack, they wrote the README, and they are the fastest answer when something in it does not work.
A reviewer reads a submission before it publishes. They check the claims on the page against what the archive actually contains, check the licence, and look for anything the maker did not mean to share — a key, a private note, an internal link.
If review finds a mistake, one of two things happens. The maker is asked to fix it, or we fix it ourselves and publish the change as a new version. Either way the template keeps its record: every version, the change note that came with it, and who uploaded it. A corrected template is visibly a corrected template.
So the maker owns the work, and we keep the order on the page. What review does and does not cover is in What review checks.
What it is not
An agent template is not a hosted service. Nothing runs on our side, and nothing keeps running when you close the page. The agent runs in your harness, on your machine, with your keys.
It is also not somebody's private setup tipped into a zip file. A pack has to be picked up by a stranger, which is a real standard — it is why the README gets its own guide, and why a template that will not explain itself does not stay up.
Where to go next
Download it, unpack it, point your harness at the folder, and read what it did.
What's inside an agent templateThe files, and the four ideas in them: a goal, a guardrail, a set of skills, and a memory.
The parts of an agent: goals, guardrails, memorySix parts, in the order they matter: instructions, skills, tools, memory, guardrails, report.
Will AI replace your job?The evidence says tasks move before jobs do, and which tasks is measurable.
Run an experiment on your own workOne task you repeat, one agent, two weeks, and a written record of what came back.