Hosted AI agents
Describe the job. AgentOS runs the agent.
Write the instructions, choose a model and give the agent its tools. AgentOS runs it on your own provider key, from the dashboard, a schedule, the API or another agent, and records every step.
The problem
The model is the easy part.
A prompt that works in a chat window is a long way from an agent that does a job every day. Somebody has to run the loop around the model, hold the credentials for every tool, start it at the right time and keep a record of what it did. That work is the same for every agent, and it is rarely what a team set out to build.
- Somewhere to run it. A server, a queue and a loop that calls the model and its tools until the job is done.
- Access to your tools. Credentials for email, the CRM and your own APIs, kept out of the prompt.
- A way to start it and check it. A trigger, a schedule, and a record of every step when something looks wrong.
How it works
From instructions to a working agent in one screen.
A hosted agent is configured in the dashboard and run by AgentOS. Your code, if any, only calls it.
- Create the agent. On Agents, press New and give it a name, a provider and a model. Or install a template from the marketplace.
- Write its instructions. The Brain tab holds the system prompt and the advanced settings: temperature, max tokens, prompt variables and an output schema. The built-in assistant can write a first prompt and suggest tools.
- Give it tools. On the Tools tab, switch on what the agent may use, and mark any tool that needs a person's approval before it runs.
- Run it, or test it first. A test run stubs out every tool that would change something, keeps the run out of your metrics, and works while the agent is still paused.
Models
Your models, on your key.
Hosted agents run on Anthropic or OpenAI, or on any OpenAI-compatible or Anthropic-compatible endpoint, such as OpenRouter, Groq or Google AI Studio. The model calls go through your workspace's own provider key, so usage is billed to your account, at your rates. Groq and Google AI Studio issue free keys without a credit card: rate-limited, but enough to try an agent.
Models that run inside your own network are supported in the self-hosted deployment, part of Enterprise.
Tools
Everything the agent may touch, in one list.
- Built-in tools. Web search, reading a web page, HTTP requests, the current time, reading and writing your files, asking a person a question, and calling another agent.
- Connected apps. Connect Gmail, Slack, GitHub, Google Sheets, HubSpot, Notion, Linear, Stripe and more once for the workspace, and their actions appear in every agent's tool picker.
- Your own APIs. Describe an HTTP endpoint as a custom tool. Authentication covers API keys, bearer tokens, OAuth2 client credentials, mutual TLS and signed requests, with secrets encrypted at rest.
- MCP servers. Connect an MCP server and import its tools, so agents call them like any other.
Pipelines
One job, several agents.
Small agents are easier to test and fix than one that does everything. A pipeline groups an orchestrator with its member agents: the orchestrator hands each part of the job to a member with the invoke_agent tool, and each member's run is linked to the orchestrator's, so the whole job reads as one tree.
Start a pipeline from its page and watch it live: which member is working, which finished, and any question an agent asks a person along the way, answered right on the timeline. The same page shows the success rate and durations across recent runs.
Ways to run
Start it from wherever the work begins.
Every run records where it came from, so the history shows what was started by a person, a schedule, your code or another agent.
- The dashboard. Press Run on the agent, or talk to it in Chat, where past conversations are kept and tool calls show as cards.
- A schedule. Daily or weekly, at a time and time zone you choose. See Scheduling.
- Your code. One HTTP call, or
invoke()in the TypeScript and Python SDKs, with streaming for chat interfaces. - Claude, Cursor or any MCP client. The AgentOS MCP server can invoke an agent, and read the runs it produced.
- Another agent. Any hosted agent can call another with
invoke_agent, which is how pipelines work.
Templates
Start from an agent that already works.
The marketplace has dozens of ready-made agents, from support triage and meeting notes to PR review and release notes, including instant starters that need no setup. Each template page says what it needs before it runs, such as a connected app or a provider key.
Installing a template copies it into your workspace as an ordinary agent, paused, so you can read its instructions and finish its setup before it does anything.
Agents you already run
Your own agents join the same record.
Not every agent should move. An agent that already runs in your own code reports its runs through the TypeScript or Python SDK, or through OpenTelemetry, and gets the same runs, traces, alerts and reports as a hosted one.
The two mix freely: an orchestrator in your code can report its own run and call hosted agents for the parts AgentOS should run, and the dashboard shows them as one tree. Schedules, tools, approvals, evals, test runs and pipelines need AgentOS to run the agent, so they are for hosted agents.
In our own workspace
Our company runs on these agents.
Our support inbox, outreach, onboarding emails and social posts all run on agents in our own AgentOS workspace.
We group agents by the job they do. Most jobs are a small pipeline: one agent hands work to the next, and a person steps in where it matters.
More in Run
Put agents to work on a schedule, with your tools and your documents.
Give one job to an agent this week.
Start with one repetitive workflow, gate the sensitive step behind your approval, and read the first run end to end.