Skip to main content
Choose an example with the framework and entry point closest to your application. These 40 projects cover Python, JavaScript, TypeScript, browser agents, web copilots, background workers, and multi-agent teams.
Repository availability does not grant a license to reuse the code. Follow each repository’s setup instructions and check its license terms before reuse.

​
Start with your own agent

Run this in your application’s repository, or give it to your coding agent:
Review the optional setup choices, authenticate in the browser, and select your project. Then connect your agent’s real entry point. The quickstart explains the complete flow; native MCP setup connects your coding agent separately. The examples use isolated local synthetic fixtures and observation-only Firedrill recipes. Their adapters invoke the application’s actual entry point. A local fixture is not a managed Firedrill Tool instance, and an observation-only recipe does not declare a quality verdict. Follow each repository’s setup instructions for its runtime and model-provider requirements.

​
Python agent frameworks

These command-based agents cover structured outputs, cited answers, stateful operations, and configuration comparisons.

​
JavaScript and TypeScript agents

Use these examples for conversational workflows, MCP-backed business actions, and agents that work with files or repositories.

​
Web apps and AG-UI copilots

These examples keep the user interface and backend agent connected. Explicit user decisions remain separate from model proposals. The included Firedrill adapters exercise the backend HTTP or AG-UI seam, not the frontend. Add a browser workflow when you need evidence of user-interface behavior.

​
Browser agents and Playwright

These JavaScript applications act through a browser rather than replacing the workflow with direct database writes. Browser evidence depends on the example’s capture configuration; not every API run has a screenshot.

​
Event-driven and background agents

Use explicit events, jobs, or scheduled ticks to explore work that extends beyond one request. Local application clocks and journals are not automatically the managed virtual clock.

​
Multi-agent teams

These Python projects explore specialist coordination, shared state, handoffs, and disagreement rather than treating every agent call as an independent answer.

​
Turn an example into your simulation

  1. Start with one small scenario and the application’s real entry point. Keep its model credentials in your own runner.
  2. Review the recipe’s inputs, selected configuration, and isolation strategy. Local synthetic dependencies do not authorize actions on real accounts.
  3. Run an observation-only simulation and inspect the recorded output and evidence. A completed execution is not automatically a pass.
  4. Add quality scoring for acceptable output variation, or exact checks for hard constraints, when they answer your question.
  5. Reuse the reviewed recipe in CI, add scenario variations, or compare model configurations.
Source examples alone do not prove live model quality, real connector delivery, hosted Tool availability, or saved cloud reports. Those depend on the accounts, configuration, and actual execution you authorize.