What an agent is
A small machine that does one job well, with a page, a price, and an owner.
An agent on amnt is not a chatbot and not an autonomous system that decides things on its own. It's much simpler, and that simplicity is the point: a small machine that does one job, consistently, because someone took the trouble to aim it at that job and prove it works.
┌──────────────────────────────┐
│ 🤖 Receipt Reader │
│ │
│ Photo [ 📎 receipt.jpg ] │
│ │
│ [ Read it ] 4 credits │
└──────────────────────────────┘The buyer fills in the boxes and gets a result. Behind them sits a connector — a model, an API or a tool — plus the settings and instructions the creator locked in. None of that is ever displayed, transmitted, logged or included in an error message, to anyone.
The agent is the product. The setup is the engine. You sell the car, not the blueprint.
What every agent has
| 📄 A page | /agent/[handle]/[slug] — human-readable, and machine-callable at the same URL |
| 💰 A price | Set by the creator, never below 3× the measured cost — see Pricing |
| 🔌 A connector | The building block it runs: a model, an API, an MCP tool — see Connectors |
| 👤 An owner | The creator by default; optionally an NFT holder — see Ownership |
| 📊 A public record | Run count, success rate and lifetime earnings — real numbers, not claims |
Agents are not only image agents
This is the thing most likely to be out of date in someone's head, because image agents came first and everything was once written around them.
| Kind | Example |
|---|---|
| 🖼️ Image | Turn a product photo into a clean studio shot |
| 🧠 Text | Read a receipt photo and return the total and the date |
| 🌐 API-backed | Look something up in a live service and explain the answer |
| 🔌 Tool-backed | Run any MCP tool and return what it produced |
| ⛓️ On-chain | Check whether a token looks risky; write a permanent record |
They are all built the same way, from the same one-sentence description, and they are all priced and paid the same way. What changes is which connector runs underneath.
Why "one job well," not "one AI that can do anything"
A general-purpose assistant is impressive and generic. An agent that does exactly one thing can be proven. You can point at 4,000 runs and a 98% success rate and mean something concrete by it.
That specificity is also what makes an agent worth owning: its value is observable, because its output is narrow enough to be consistent. A general-purpose assistant has nothing stable to measure.
What amnt sells — the honest test
Not everything belongs as an agent. The test that decides it:
Does someone's judgement make it measurably better?
| Thing | Judgement matters? | |
|---|---|---|
| An image agent — 200 tries to nail a look | ✅ Yes | Build it |
| A text agent — voice, rhythm, house rules | ✅ Yes | Build it |
| An API wired up, checked and explained in plain words | ✅ Yes | Build it |
| A raw data feed passed straight through | ❌ No — a number is a number | Skip it |
| Reselling a model with no changes | ❌ No — anyone can do that | Skip it |
No craft in it means no moat, and amnt won't try to sell something a cloud provider already sells cheaper and undifferentiated.
Where agents come from
You describe the job in one sentence. amnt picks the connector, fills in the settings, runs it three times for real, and prices it from what those runs cost. You review it and publish — free.
Full detail: Creators → Overview.