A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression

2026-04-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

The TACO (Terminal Agent Compression) framework addresses the quadratic token cost growth in long-horizon, multi-turn terminal-centric agentic tasks by automatically discovering and refining compression rules. This plug-and-play, self-evolving framework is designed to work with existing terminal agents and overcomes the generalization difficulties of heuristic or fixed-prompt compression methods in heterogeneous terminal environments. Experiments across TerminalBench (TB 1.0, TB 2.0), SWE-Bench Lite, CompileBench, DevEval, and CRUST-Bench demonstrate TACO's effectiveness. When integrated with MiniMax-2.5, TACO improves performance on most benchmarks while reducing token overhead by approximately 10%. On TerminalBench, it yields consistent performance gains of 1%-4% across strong agentic models and further boosts accuracy by 2%-3% within the same token budget.

Key takeaway

For NLP Engineers developing long-horizon terminal agents, TACO offers a significant advantage by mitigating cumulative token costs and improving performance. You should consider integrating this self-evolving compression framework to enhance the efficiency and accuracy of your agentic models, especially when working with diverse terminal environments. This can lead to better resource utilization and more robust agent behavior.

Key insights

TACO is a self-evolving framework that compresses observational context for terminal agents, improving performance and reducing token costs.

Principles

• Self-evolving compression adapts to diverse terminal environments.
• Task-aware compression enhances agent performance and efficiency.

Method

TACO automatically discovers and refines compression rules from interaction trajectories, integrating as a plug-and-play component for existing terminal agents.

In practice

• Apply TACO to reduce token overhead in long-horizon agent tasks.
• Integrate TACO with existing agent frameworks for performance gains.

Topics

• Terminal Agents
• Observational Context Compression
• Self-Evolving Frameworks
• Long-Horizon Reasoning
• TerminalBench

Best for: NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression — TACO enables self-evolving, task-aware compression for terminal agents, significantly reducing token costs and improving performance.
• LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents — Synthetic, verifiable environments offer a scalable solution for training language agents in complex terminal tasks.
• 🥇Top AI Papers of the Week — AI advancements focus on novel computing, efficient memory, robust agents, and specialized models.
• Signal-Driven Observation for Long-Horizon Web Agents — Decoupling web agent observation from action frequency via signal-driven DOM filtering prevents context degradation.
• Notation Matters: A Benchmark Study of Token-Optimized Formats in Agentic AI Systems — Token-optimized formats like TRON can significantly reduce LLM token consumption in agentic systems with minimal accuracy loss.
• GUIDE: Guided Updates for In-context Decision Evolution in LLM-Driven Spacecraft Operations — GUIDE enables LLMs to adapt in real-time by evolving natural-language decision rules without weight updates.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.