Skip to main content

​
Constructor

new Agent(options: AgentOptions)

​
Options

AgentOptions includes:
  • llm, storage, embedder
  • planPrompt, reasonPrompt, actPrompt
  • k (default: 3), maxSteps (default: 30)
  • seedTrajectories, onStep
  • curationThreshold (default: 0.3), curationMinRetrievals (default: 5)
  • verifyTrajectory

​
Core Methods

await agent.init();
await agent.train(env, goal);
await agent.run(env, goal);
await agent.trainBatch(envFactory, goals);
await agent.runBatch(envFactory, goals);
agent.getDatabase();
agent.getStats();

​
Notes

  • init() must be called before train() or run().
  • train stores successful trajectories (subject to verifyTrajectory).
  • run never writes new trajectories.
  • getDatabase() exposes the underlying TrajectoryDatabase.

​
Stats Shape

getStats() returns:
  • totalTrajectories
  • successfulTrajectories
  • successRate