Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.agno.com/llms.txt

Use this file to discover all available pages before exploring further.

The template ships a Claude Code prompt at docs/create-new-agent.md that handles the mechanical parts of adding an agent: creating the file in agents/, registering it in app/main.py, restarting the container, and running a smoke test.

Run the prompt

Open Claude Code in the repo and paste:
Run docs/create-new-agent.md
Claude will ask:
  1. What should the agent do? A one-line description. “Summarize GitHub issues for a given repo,” “Triage incoming Slack messages,” etc.
  2. What tools does it need? Pick from Agno’s 100+ toolkits or describe a custom tool.
  3. Any specific instructions? Tone, format, escalation rules, anything you’d put in the system prompt.
Claude generates agents/<your_agent>.py, adds the import and registration to app/main.py, runs docker compose restart, and hits the new agent’s endpoint to confirm it responds. A simple agent (one toolkit, no custom tools) takes 5–10 minutes. A bespoke agent with custom tools takes longer.

Test in the AgentOS UI

Open os.agno.com, select your new agent in the sidebar, and try realistic prompts. Run it through:
  • The golden path. What you built it for.
  • Edge cases. Unusual inputs, partial information, ambiguous questions.
  • Adversarial inputs. Prompt injection, out-of-distribution requests, attempts to make it do something it shouldn’t.
Click Sessions and Traces in the sidebar. Every run is recorded with full message history, tool calls, and timing. This is what powers the iteration loop on the next page.
Spend 5–20 minutes in the UI before moving on. The traces you generate here become the input data for the improvement loop.

What just happened

ActionWhere it landed
Agent definitionagents/<your_agent>.py
Registrationapp/main.py
Endpointhttp://localhost:8000/agents/<your_agent>
Sessions and traces from your test runsPostgres, viewable in the AgentOS UI

Next

Improve the agent →