SKILL.state architecture replaces conversation history with mutable execution state for long-term agents

Problem: unbounded history

Large language models (LLMs) that act as autonomous agents encounter a structural issue when tasks extend over time. Existing runtimes store every observation, action, and intermediate reasoning trace in a conversation history that grows indefinitely. Adding only new entries creates two main failures: latency that worsens with each step and “context poisoning” where accumulated noise diverts the model from the intended path.

Solution: explicit mutable execution state

The authors introduce SKILL.state, a runtime architecture that swaps the conversation history for an explicit, structured, mutable execution state. At each step the model receives three inputs only: an immutable skill specification, the current execution state, and the latest observation. Intermediate reasoning traces are discarded immediately after they produce a verified state update, preventing the prompt from expanding with task length.

Results: fewer tokens, higher accuracy

The architecture was evaluated across multiple datasets, models, and runtime environments. Reported outcomes show improved task accuracy together with a marked reduction in cumulative token consumption. The central claim is that an explicit execution state provides an efficient, architecture-agnostic abstraction for long-horizon agent skills.

Limitations and context: pre-print, not peer-reviewed

The paper appears on arXiv as a pre-print and has not undergone formal peer review. Code and demos are publicly available, enabling independent replication, but the metrics come from the authors without external validation. Moreover, the comparison was made against a basic add-only approach; it remains unclear how SKILL.state performs relative to more advanced context-management techniques such as hierarchical summarization or selective memory retrieval.