Agents
An agent is an AI assistant that does a job for you using the systems you have connected. Ask it “where is order NW-10428?” and it looks the order up, checks the shipment and answers in a sentence. Ask it for a refund and it can issue one, or stop and ask a person first.
On this page: the two kinds of agent, how to create one, and where to go next.
Two ways to build an agent
| Kind | Pick it when | What you do |
|---|---|---|
| Autonomous | The work changes from one request to the next, like answering customers | Write instructions in plain words and say what it may use. The AI works out the steps itself. See Autonomous agents. |
| Workflow | The same steps must run the same way every time, like a nightly report | Draw the steps on a canvas and connect them. See the visual builder. |
Both kinds use the same pieces: the tools and APIs you connected, the AI models you connected, memory, checks by a second model, schedules and webhooks. You choose per agent and give nothing up either way.
Which AI model does the work is set separately from what the agent does, so you can switch models at any time. An autonomous agent picks its work mode and a model for each part on its edit page. A workflow agent picks its models on its Execution tab; see Roles and strategies.
Create an agent
- Open Agents in the sidebar and click Create agent.
- At the top of the builder, pick Autonomous or Workflow.
- For an autonomous agent, write its personality and instructions, pick a work mode and its models, then tick what it may use under Capabilities. For a workflow, the canvas opens with a simple Input, AI call, Output flow for you to extend.
- Save, then activate it. An agent answers only while it is active.
From the agents list you can also activate, pause, duplicate or delete an agent. Deleting an agent also deletes its channels, so its web chat, website bubble and messaging channels stop answering. Import brings in an agent saved as a file (Import from JSON) or an agent that runs somewhere else (Import external A2A agent; A2A is a standard way for agents to talk to each other).
A second opinion before it answers
An autonomous agent can have its answer checked before anyone sees it. One or more checker models each try to find something wrong with the answer. If one does, the agent fixes the answer and tries again. Use a checker from a different AI vendor than the main model, so one vendor’s mistake does not slip through. Set this up on the agent’s Overview tab; details in Autonomous agents.
An agent’s status
| Status | What it means |
|---|---|
| Draft | Still being built; nobody can run it yet |
| Active | Live and answering |
| Inactive | Paused; requests are turned away |
| Error | Something in its setup needs fixing |
Where to next
- Autonomous agents: checks, rules, skills, memory
- Visual builder: draw a workflow step by step
- Step types: every step the builder offers
- Scheduling and Webhooks: run an agent on a timer, or tell another system when it finishes
- Versioning: look back at changes and undo them
For developers
List and run agents from the terminal with the CLI:
$ npx @almyty/agents list
$ npx @almyty/agents run invoice-pipeline --input '{"file_url": "https://example.com/invoice.pdf"}'Autonomous runs stream their steps with --watch and take safety limits like --max-steps and --max-cost-cents. Full command reference in the Agents CLI.
Every agent, whichever mode, also answers on the OpenAI-compatible API, so an existing OpenAI SDK can call it by changing the base URL.
The status values in the API are draft, active, inactive and error.