Stateful Tracking in Agent Loops

Maintain explicit state of what the agent already knows so it does not re-fetch information that was retrieved in a prior turn unless it has changed.

How It Works

A lightweight state object tracks file contents, directory listings, tool results, and user preferences across turns. Before making a tool call, the agent checks its state to see if the information is already available and still fresh.

What to Track

• file contents and modification times
• directory listings
• previous tool outputs
• user preferences and constraints
• the current plan and completed steps

Benefits

• eliminates redundant reads
• reduces prompt size by avoiding re-inclusion of unchanged data
• prevents the agent from repeating work

Anti-patterns

• Re-reading a file after every edit to "confirm" the change.
• Sending the full directory tree on every turn instead of caching it.
• Re-fetching tool definitions or repository context that has not changed.

See Also

• Operation Batching — state tracking enables more effective batching by knowing what is already cached.
• Incremental Plan Updates — plan state is a key component of agent state.
• Diff-Based Tool Output — tracking file state enables diff computation.
• Structured State Extraction — takes this further by discarding the transcript entirely in favor of a schema-validated state object.
• Event-Driven Waiting — tracking what is already in flight prevents a re-dispatch after a spurious timeout.

Tools

• LangGraph — framework for building stateful, multi-actor agent loops with explicit state management.
• Aider — tracks repository state to avoid redundant file reads and context rebuilds.