Concepts
What is an AI agent?
An AI agent is a program that uses a language model to reason about a task, decide what to do, and then take actions using tools — not only produce text. The distinction that matters is the acting part: a model that writes you an email draft is doing one job, and a program that writes the draft and then sends it, and records that it sent it, is doing a different one.
I build a hosted agent platform, so treat the framing as interested. Everything below that is a number comes from the product itself.
How it differs from the two things people usually mean
| What it does | Where it struggles | |
|---|---|---|
| Chatbot | Answers in text, turn by turn | It cannot do anything. Every action is left to the person reading |
| RPA | Repeats pre-recorded steps exactly | Variation. A field moves and the script breaks, loudly if you are lucky |
| AI agent | Reasons over each case, then calls tools to act | Being confidently wrong. It rarely breaks; it reports success |
That last row is the honest cost of the approach, and it is worth knowing before you build anything. Source: a run of mine called an API twice, got 403 both times while the tool result said "ok": true, and proceeded as though it had the data — telling the customer it had checked her orders. It finished COMPLETED.
What an agent here can actually do
- Reason over an input and return an answer, optionally as structured JSON other steps can read.
- Call any REST API, send email, fire a webhook, post to Slack, Discord, Telegram or Google Chat, create a Notion page, and read or create Google Calendar events — ten tools in total.
- Hand off to other agents in a workflow, in sequence, in parallel, on a condition, fanned out, in a loop, or as a dependency graph.
- Run without you, on a cron schedule or when an inbound webhook fires.
- Record every step: what it received, what it returned, its tokens, its cost and its duration.
What one costs to run
Agents are billed in tokens, so the cost follows the size of the job rather than a per-task fee. A worked support-ticket example comes to $0.0144; a real logged run that also escalated to a human $0.0204; a run that fails on its first tool call $0.0024. The dearest agents shipped here reach about 11 cents a run.
Source: the worked example shows two of those costed step by step, no signup needed. That spread — a 46× range between the cheapest and dearest run — is why a single "cost per agent" figure is not worth quoting.
The parts people underestimate
Tools are the whole point, and the whole risk
An agent with no tools is a text generator. An agent with tools can email your customers. Every tool here is opt-in per agent and capped per day — 100 emails, 1,000 HTTP calls — because an agent on a schedule runs with nobody watching.
A status is not a result
The single most useful habit is reading the step rows rather than the run status. A run status only knows whether something raised an error, and the expensive failures do not raise anything.
Judgment is the reason to use one
If the steps are identical every time, a deterministic automation is cheaper, faster and easier to trust. Reach for an agent when the input varies enough that rules would be brittle.
Next
- Build one — see Create and configure an agent.
- Compare the approaches properly — see AI agents vs automation tools.