TACO - An open-source, self-evolving observation compression framework for edge-level intelligent agents from universities such as Beihang University.

What is TACO?

TACO is an open-source, plug-and-play, training-free observation compression framework for self-evolving terminal agents, developed by the University of Manchester, Beijing University of Aeronautics and Astronautics, Hong Kong University of Science and Technology, and the MAP team. The framework addresses context bloat caused by accumulated shell output noise in multi-turn command-line tasks by automatically discovering and reusing compression rules from interaction trajectories, preserving key errors while filtering redundant logs. The framework is plug-and-play and training-free, and is already integrated into Harbor's terminus-2. On benchmarks such as TerminalBench, it delivers 1%-4% accuracy improvements to models like DeepSeek-V3.2 and Qwen3, significantly reducing token consumption for long-duration tasks.

TACO's main functions

• Discovery of self-evolutionary rulesThe framework automatically scans the raw shell output from multiple rounds of terminal interaction trajectories, identifies redundant patterns, and generates candidate compression rules, eliminating the need for manually writing fixed prompts or heuristic strategies.
• Online rule refinement and repairIteratively adjust rule boundaries based on real-time feedback from task execution, fix issues of over-compression or omission of key signals, and ensure that error messages and status feedback are not mistakenly deleted.
• Global rule pool and cross-task migrationMaintain a persistent global rule knowledge base so that new tasks can directly load and reuse rules validated in previous sessions, enabling experience accumulation across repositories and command environments.
• Plug and play, training-free integrationIt can be directly embedded into existing terminal agents (such as terminus-2 in the Harbor framework) as a plugin, and can be enabled via command line parameters without modifying the model architecture or making any fine-tuning.
• Token efficiency and performance optimizedWhile filtering out low-value terminal noise, it retains key environmental feedback, making token consumption linear and controllable for long-term, multi-round tasks, and improving accuracy by 1%-4% on benchmarks such as TerminalBench.

TACO's technical principles

• Problem BackgroundThe terminal agent completely fills the context with the original shell output in multiple rounds of tasks, causing low-value noise such as system logs and progress bars to expand twice with each round, drowning out key error signals and drastically increasing token costs.
• Three-module architectureThe framework consists of a rule discoverer, a rule refiner, and a global rule pool. The discoverer monitors the output stream and automatically generates structured compression rules for excessively long and uncovered content; the refiner iteratively repairs rule boundaries based on task feedback to prevent accidental deletion of critical information; and the global rule pool persists verified rules for direct loading and reuse by new tasks.
• Online operation mechanismIn each round of interaction, compression is prioritized by matching the global rule pool. If a long output that is not covered is encountered, a new rule discovery is triggered, which is then incorporated into the global pool after verification by the current task. The entire process is driven by an externally planned LLM, requiring no training data or model fine-tuning, allowing the compression strategy to evolve dynamically with the terminal environment.

How to use TACO

• Environmental installationAfter cloning the GitHub repository, execute the following command in the project root directory: pip install -e . After completing the dependency installation, TACO is directly available as a terminus-2 component of the Harbor framework.
• Quick StartRun the Harbor command and specify the terminus-2 agent, passing in the model information and TACO-specific parameters to start the process. scripts/run_taco_example.sh Provides directly modifiable template scripts.
• Core parameter configuration:pass --ak Prefix input enable_compress=True Turn on the main compression switch.enable_self_evo=True Activate online rule evolution and configure simultaneously compress_base_url,compress_api_key and compress_model_name Points to any OpenAI-compatible planned LLM endpoint.
• Common mode selectionFor a complete experience, compression and self-evolution must be enabled simultaneously, and an external LLM must be configured; if fixed rules are required for ablation experiments, add the following: freeze_rules=TrueIf only single-task local evolution is used without inheriting the global rule pool, then add... disable_global_evo=True.
• Operation control:use --ak max_turns=200 Set the maximum number of rounds per task using parameters such as [parameter name], and then [do something]. model_info Pass in JSON to configure LiteLLM's input and output token limits, ensuring that long-running tasks are executed within a controllable cost.

TACO's core advantages

• Plug and play, zero trainingNo need to fine-tune the main backbone model or modify the Agent's underlying architecture; compression and self-evolution can be enabled directly via command-line parameters.
• Self-evolving cross-task migrationThe global rule pool allows the Agent to continuously accumulate and compress experience in long-term, multi-round tasks, and new tasks directly inherit old knowledge.
• A win-win situation in terms of performance and costIt delivers a 1%-4% accuracy improvement on models such as MiniMax-M2.5, DeepSeek-V3.2, and Qwen3-Coder-480B on TerminalBench, and an additional 2%-3% improvement with the same token budget.
• Strong generalization and low consumptionIt maintains or improves success rates on SWE-Bench Lite, CompileBench, DevEval, and CRUST-Bench, while significantly reducing total token consumption.

TACO's project address

• GitHub repository: https://github.com/multimodal-art-projection/TACO
• arXiv technical paperhttp://arxiv.org/abs/2604.19572

Comparison of TACO's similar competing products

TACO application scenarios

• Long-term software engineering agentSuppress log and terminal output explosion and maintain clear context during multi-round code debugging and compilation testing in SWE-Bench.
• Automated operation and maintenance and deploymentIt handles a large amount of redundant system status and process information returned by the shell, improving the decision-making stability of the DevOps Agent.
• Code review and test analysisFilter out irrelevant compilation warnings and test pass information, accurately retaining key errors and code differences.
• Academic Research Reproduction and EvaluationAs a terminus-2 plugin for the Harbor framework, it is used for token efficiency evaluation and long-range capability benchmarking of terminal agents.