Basics
Get started in 3 steps
AgentMesh lets you build AI agents that do real work — answer questions, call your systems, and run on their own. Here is the fastest path to seeing one work. It takes about two minutes on the free plan, which needs no card.
The three steps
- On your Dashboard, click “✨ Draft a support reply with a starter agent”. This creates a ready-to-run agent and runs it on a sample support ticket.
- Watch the run page — the agent’s answer appears when it finishes (usually a few seconds).
- That’s your first successful run. Now open the agent to change its instructions, or build your own from scratch.
Prefer to build your own? Click “Create agent”, give it a name and instructions, set it to Active, and run it.
What a first run actually costs
Runs are billed in tokens, so the cost depends on the job rather than on a per-task fee. These are real numbers from runs we published, not estimates:
| A run of this kind | Time | Cost |
|---|---|---|
| One support ticket triaged and answered (worked example) | 8.4 seconds | $0.0144 |
| The same job, escalated to a human as well | 28.5 seconds | $0.0204 |
| A failed run, where the API was down | 4.1 seconds | $0.0024 |
Every figure above is checkable, and the two sources are not the same kind of thing. Source: row two is mine — I ran it on my own account and published the raw execution record exactly as the API returned it, with one recipient address redacted. Rows one and three are the worked example on see a real run — a representative job rather than a customer’s, because that page is public and a real one would expose somebody’s input.
The costs there are not estimates: they are computed from the token counts by the same pricing code the product bills with, which is why they run to four decimal places instead of being rounded to something tidier.
The same agent on two tickets can differ by 1.4× on cost and 3.4× on time, which is why a single “cost per run” figure is not worth quoting.
If the first run fails
It finished but the answer is empty
Open the run and read the step rows. Every step records what it received and what it returned, so an empty answer usually points at one step that got nothing to work with.
A tool call shows HTTP 403 or failed
The agent reached something it is not allowed to use, or an integration is not connected yet. A refused call now shows its status next to the tool name rather than passing silently.
Nothing ran at all
Agents are inactive by default so nothing runs by accident. Open the agent and set it to Active.
What to do next
- Change the starter agent’s instructions and run it again — that is the fastest way to see how the prompt drives the behaviour.
- Give it a tool so it can act: call an API, send an email, post to Slack.
- Put it on a schedule or behind a webhook so it runs without you.