AI Agents · Lesson

Trajectory-Based Self-Improvement

Learning from successful and failed action sequences to refine future behavior.

What Is a Trajectory?

A trajectory is the complete sequence of states and actions an agent took from start to finish for a given task. It records not just the final answer but how the agent got there: which tools were called, in what order, with what parameters, and what intermediate results were observed.

Recording a Trajectory

Wrap each agent action in a recorder that captures state before and after. A state includes: the current goal, memory contents, and recent observations. An action includes: tool name, parameters, and result.

from dataclasses import dataclass, field
from datetime import datetime
from typing import Any

@dataclass
class TrajectoryStep:
    step_index: int
    state_summary: str    # short description of world state
    action_name: str      # tool or reasoning step name
    action_params: dict
    result: Any
    timestamp: str = ''

    def __post_init__(self):
        if not self.timestamp:
            self.timestamp = datetime.utcnow().isoformat()

@dataclass
class Trajectory:
    trajectory_id: str
    task: str
    steps: list = field(default_factory=list)
    outcome: str = 'unknown'   # 'success', 'failure', 'partial'
    final_score: float = 0.0

    def add_step(self, step: TrajectoryStep):
        self.steps.append(step)

    def mark_success(self, score: float = 1.0):
        self.outcome = 'success'
        self.final_score = score

    def mark_failure(self, reason: str = ''):
        self.outcome = 'failure'
        self.final_score = 0.0

All lessons in this course

← Back to AI Agents