State across context windows

Definition

Techniques for carrying work across multiple context windows in long-running agentic tasks, so progress survives beyond a single window's token budget.

Key points

• Use git for checkpoints — commit state so later windows can resume from a known point.
• First window sets up the framework (tests, scripts, scaffolding); later windows iterate on a todo list against that framework.
• Use the memory tool plus context awareness to track what remains.
• Isolated subagent context (Subagent decision criteria) also limits how much state any one window must hold, mitigating Context rot.
• Related mechanisms: session --resume <id> to continue and --fork-session to branch without modifying the original.

Why it matters for the exam

• Code-generation and multi-agent scenarios test durable multi-window strategy: git checkpoints, framework-first, todo-driven iteration — not cramming everything into one window.

Common gotchas

• Progressive summarisation compounds detail loss; prefer durable external state (git, memory tool, todo files) over re-summarising context.
• Keep critical constraints in the system prompt so they survive across windows.

See also

• Subagent decision criteria
• Context Rot
• Agentic loop

Sources

• netlight-study-guide