Skill.State

SKILL.state

**SKILL.state** is a long-horizon agent skill pattern for Claude that replaces append-only chat history with an explicit structured execution state. It is designed for multi-step workflows where the agent must remember only what is relevant for future actions, not the entire transcript. In the paper’s benchmarks, the key measurements are prompt size, total token cost, task accuracy/success rate, noise robustness, and recovery steps across horizons such as 10, 25, 50, 100, and 200.

Overview

Traditional agent loops keep adding observations, reasoning, and tool outputs to the prompt, which makes the context grow and can degrade performance over time. SKILL.state instead keeps three things at each step: the immutable task spec, the current state, and the latest observation. The model proposes a state patch and next action, the runtime validates the patch, applies it, and then discards the reasoning trace.

Why use it

This pattern is useful when the task is:

• long-running,
• stateful,
• tool-heavy,
• dependent on active entities, hypotheses, files, or inventories,
• sensitive to noisy or outdated history.

It is especially useful when repeated history reconstruction causes errors or token growth.

Execution flow

1. Start with a task specification and a structured state.
2. Provide Claude with only the current state and the latest observation.
3. Claude returns:
   • a state_patch,
   • an action.
4. Validate the patch against the schema.
5. Merge the patch into the current state.
6. Execute the action and continue with the next observation.

This keeps execution bounded and avoids relying on stale transcript content.

State patch rules

The state patch should be minimal and only contain changes needed for future execution.

Rules

• Update only changed fields.
• Use null to delete a key.
• Do not rewrite the entire state unless necessary.
• Keep the state compact and operational.
• Prefer structured values over free text.
• Sto