LangGraph Autonomous Task Agent (Sidekick)

Static-answer assistants break down on tasks that require browsing, extraction, intermediate memory, and stopping criteria.

The goal here was to build a more execution-oriented assistant that can make progress through a sequence of tool-backed steps.

The architecture uses a LangGraph StateGraph so agent state, tool usage, and evaluation decisions remain explicit rather than hidden in prompt history.

This makes retry behavior, memory updates, and completion checks much easier to reason about.

Playwright handles browser automation, while SQLite-backed checkpoints give the agent stable memory between iterations.

A tool layer wraps browsing, retrieval, and notifications so the orchestrator can route the right capability at the right step.

The project demonstrates a practical pattern for stateful assistants that need to act on the web instead of only reasoning over text.

Task agents benefit from explicit graph structure because debugging hidden control flow inside prompts quickly becomes unmanageable.