Agentic System as Compressor: Quantifying System Intelligence in Bits

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Zihan Qin, Hongrui Zhang
• Venue: arXiv: cs.AI
• Published: 2026-06-24
• Collected: 2026-06-25
• URL: http://arxiv.org/abs/2606.25960v1
• PDF: https://arxiv.org/pdf/2606.25960v1
• DOI: UNKNOWN

Plain-Language Brief

The authors propose a method to measure the intelligence of an agentic system by quantifying its ability to compress a target object under a fixed task distribution, interface, and compute budget. This is done using arithmetic coding, seed coding, and a fallback in five different settings: reversed text, chess moves, protein sequences, retrieval-augmented question answering, and semantic story compression.

What the Document Says

• The paper discusses the transition of large language models from isolated predictors to agentic systems.
• The authors adopt an analytical viewpoint based on the principle 'compression is intelligence'.
• They propose a method to measure the strength of an agentic system by quantifying its ability to compress a target object under fixed task distribution, interface, and compute budget.
• This measurement is operationalized using arithmetic coding, seed coding, and a fallback.
• The authors evaluate this method in five settings: reversed text, chess moves, protein sequences, retrieval-augmented question answering, and semantic story compression.
• In all settings, the use of agentic components reduces codelength.

Key Concepts Mentioned

• Large language models
• Agentic systems
• Compression is intelligence
• Arithmetic coding
• Seed coding
• Fallback
• Task distribution
• Interface
• Compute budget
• Codelength

Claims Made by the Document

• A stronger agentic system lets a target object be reconstructed with fewer bits.
• The proposed method can analyze how components, observers, and budgets change residual uncertainty.

Evidence Used

• Five small, controlled experiments covering typical component types of real agentic systems
• Reduction in codelength in all five settings

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

• The authors acknowledge that their experimental settings are small and controlled, which might limit the generalizability of their findings.

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

• How does this method compare to other methods for evaluating agentic systems?
• What are the implications of this method in real-world scenarios?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper focuses on a specific method for measuring the intelligence of an agentic system. However, it's important to note that this is just one approach and there might be others.
• While the experiments show promising results, more research would be needed to fully understand the applicability and limitations of this method in real-world scenarios.

Raw Abstract / Excerpt

Large language models are turning from isolated predictors into agentic systems: they call tools, retrieve evidence, obey environment constraints, use verifiers, and complete tasks through search and multi-turn interaction. We adopts an analytical viewpoint based on 'compression is intelligence': under a fixed task distribution, interface, and compute budget, a stronger agentic system lets a target object be reconstructed with fewer bits. We operationalize the measure with arithmetic coding, seed coding, and a fallback, and evaluate it in five settings: reversed text, chess moves, protein sequences, retrieval-augmented question answering, and semantic story compression; in all of them agentic components reduce codelength. These small, controlled experiments cover component types typical of real agentic systems, show that codelength can analyze how components, observers, and budgets change residual uncertainty, and offer guidance for evaluating real agent systems.