XYZ-AI-Lab/AxisAgentic - 仓库详情

AI 仓库情报摘要

FR-AI / ANALYSIS

为什么值得关注

该框架通过仅追加轨迹记录重建模型在各个阶段可见的上下文，将智能体执行与学习过程独特地连接起来，从而支持回放、评估和状态忠实的监督微调。

适合谁使用

开发长程AI智能体的研究人员
构建包含工具编排的模块化智能体系统的开发者
需要可靠轨迹收集以进行监督微调的团队
寻求可复现智能体评估的基准测试人员

典型使用场景

构建和评估具有上下文管理与恢复功能的多轮搜索智能体
收集并导出状态忠实轨迹，用于模型的监督微调
回放和调试智能体运行，以理解模型在每个步骤的行为
使用一致的评估框架对智能体进行标准任务基准测试

项目优势

仅追加轨迹设计保证在任何阶段都能忠实重建模型可见的上下文
完全模块化架构允许交换模型客户端、工具、编排器、数据集和评估器
内置长程鲁棒性支持：上下文预算、压缩、回滚、重试、恢复、自我验证
提供带有运行时可见性标记和来源信息的SFT导出器，支持轨迹选择

使用前须知

需要 Python 3.12+ 和 OpenAI 兼容端点，限制了在旧版或非标准环境中的部署
不包含模型权重，用户必须自行提供模型和端点
参考配方目前集中于搜索应用（Web Search、WideSearch），通用或编码智能体配方可能不够成熟

README 快速开始

English · 简体中文

XYZ AI Lab ·
Technical Report

AxisAgentic is an extensible runtime for long-horizon AI agents. It also collects the trajectories produced during execution. The runtime works with OpenAI-compatible endpoints and pluggable local model clients, and handles multi-turn execution, tool orchestration, context management, recovery, and benchmark evaluation. Each trace preserves the state visible to the model, so the same record can support recovery and replay, benchmark evaluation, or filtered SFT export.

Web Search and WideSearch are the current reference recipes. The same extension points can support domain, general-purpose, and coding agents. This repository does not include model weights.

✨ Core capabilities

Append-only traces preserve runtime events and reconstruct the context visible to the model at any stage.
Model clients, tools, orchestrators, datasets, evaluators, reward functions, and recipe policies are replaceable.
Context budgets, compaction, rollback, retries, recovery, self-verification, and tool limits support long runs.
Each task records traces, token and timing metrics, evaluation artifacts, and provenance for trajectory selection.
SFT exporters replay runtime visibility markers, while rollout interfaces connect execution to external training systems.
Strict YAML schemas and portable path schemes keep runs reproducible across environments.

🔁 From execution to learning

Every task writes an append-only trace. The trace is the common source for replay, evaluation, and trajectory collection. Selected trajectories can then be exported for external training.

Runtime markers record rollback, context compaction, and discard-all events. Replaying them reconstructs what the model saw at a given stage. Trace inspection and SFT export use the same rules, so supervised examples exclude hidden history and rolled-back actions.

Recipe exporters emit Swift Agent and related training formats with the source trace, task status, and optional metadata. The external training pipeline owns final correctness filters, loss masks, and optimization. AxisAgentic supplies the replay and export boundary so inference and training use the same interaction history.

🦅 Flagship reference: XYZ-Aquila

XYZ-Aquila is a search system built with AxisAgentic. Its recipe combi

项目描述

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

AxisAgentic