Agent State Management

Introduction

State = everything the agent knows mid-run: plan, step results, scratchpad, memories, tool call history. Mismanaging state is the #1 source of subtle agent bugs.

Beginner analogy: a chef's order ticket — must reflect exactly what's been cooked vs pending; lose it and chaos.

Understanding the topic

Core concepts:

• Use an explicit state object (LangGraph, Pydantic state).
• Make state serialisable — for pause/resume and traces.
• Persist state to a DB for long-running agents.
• Idempotent steps make retries safe.
• Diff state between steps to find unexpected mutations.

Syntax reference

Visual workflow / architecture:

state = {
  goal: "…",
  plan: ["a","b","c"],
  current_step: 1,
  results: { a: {...} },
  memory_ids: [...]
}
─► serialise → save → restore later

Real-world use

LangGraph centres on a typed state; Inngest persists agent state across hours; Vercel AI SDK exposes useChat state primitives.

Best practices

• Type your state — TypeScript or Pydantic.
• Persist after every step for crash-safety.

Common mistakes

• Mutating shared state across parallel branches without locks.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why use a typed state object?
• Q2. What does it mean for a step to be idempotent?
• Q3. Scenario: your agent crashes mid-run and restarts from scratch. Fix?