XYZ-AI-Lab/AxisAgentic

AxisAgentic: An Extensible Runtime and Trajectory-Collection Framework for Long-Horizon Agents.

What it solves

AxisAgentic provides a standardized runtime for long-horizon AI agents, addressing the difficulty of managing complex, multi-turn interactions and collecting high-quality training data from those interactions. It ensures that the data collected for supervised fine-tuning (SFT) exactly matches what the model actually saw during execution, avoiding discrepancies between inference and training.

How it works

The framework uses an append-only trace system to record every runtime event, including tool use, context compaction, and rollbacks. This allows the system to reconstruct the exact state visible to the model at any point in time. It supports pluggable components—such as model clients, tools, and reward functions—and handles the operational overhead of long runs through context budgets, retries, and self-verification.

Who it’s for

It is designed for developers and researchers building complex AI agents that require long-term memory, tool orchestration, and a pipeline for improving models via trajectory collection and SFT export.

Highlights

  • State-Faithful Traces: Records runtime markers to ensure SFT exports exclude hidden history or rolled-back actions.
  • Extensible Architecture: Replaceable model clients, orchestrators, and recipe policies.
  • Long-Horizon Support: Built-in mechanisms for context budget management, rollback, and recovery.
  • Execution-to-Learning Loop: Seamlessly converts runtime trajectories into training formats for external systems.
  • Reference Implementations: Includes recipes for Web Search and WideSearch, as demonstrated by the XYZ-Aquila system.

Related

  • Project
    ApodexAI/FrontierAgent

    FrontierAgent is an open‑source, terminal‑based runtime that lets you run autonomous ReAct agents or parallel Agent‑Team workflows. It provides a sandboxed filesystem, live TUI task board, asynchronous user interventions, and a built‑in benchmark harness for research‑grade long‑horizon tasks. Install with `uv`, point it at any OpenAI‑compatible model endpoint (including the free Apodex‑1.1 service), and run either `--mode react` for a single agent or `--mode agent_team` for coordinated parallel agents. The project is Apache‑2.0 licensed.

  • Project
    verl-project/uni-agent

    Uni‑Agent is an open‑source framework that unifies agent harnesses, sandbox execution, and RL training, enabling thousands of concurrent long‑horizon LLM agents with reproducible recipes and benchmarked results.

  • Project
    agentspan-ai/agentspan

    A durable runtime for AI agents that prevents execution loss during process crashes and supports long-running, event-driven workflows with human-in-the-loop approvals.

  • Project
    AgentEra/Agently

    Agently is a Python framework for building reliable AI services. It provides structured request/response contracts, instant streaming of partial results, observable actions (tool calls), versioned skills, and a signal‑driven workflow engine (TriggerFlow). The library abstracts model providers, supports model pools, and includes built‑in actions for Python, shell, web browsing, SQLite, etc. Designed for production‑grade assistants, internal copilots, and AI‑backed APIs that need debuggability, restart‑safe workflows, and strict output validation.

  • Project
    agentscope-ai/Trinity-RFT

    A general-purpose framework for reinforcement fine-tuning of LLMs that decouples exploration, training, and data buffering to support diverse RL algorithms and agentic workflows.