Headlong：面向持久代理的开源微框架

Laude Institute 于 2026 年 8 月发布了开源人工智能代理微框架 Headlong，相关论文由 Laude Institute 与 MIT 联合发表于同期 1。该框架的核心特点是“持久代理”，允许智能体在没有外部输入的情况下，通过自我引导循环持续思考 1。在技术实现上，Headlong 的核心代码少于 1 万行 Bash，其中 bin/ 和 thinkers/ 目录下的代码共计 9900 行 1。此外，该框架支持多用户共享单一思维流、分层上下文压缩以及 DAG 轨迹格式 1。

在实际运行中，Headlong 展现了自主修复代码 bug 的能力 1。例如，智能体 Audel 在 8 月 5 日自主构建并修复了 recall 进程，从检查到验证修复耗时 48 分钟，对应 commit 80cbb1e 1。同时，Audel 自主发现并修复了 guard 中的 bug，对应 commit da31e98 1。在思维合并方面，Headlong 的 shellm 递归子运行在前两天合并回 Audel 思维 64 次，此后 12 天内合并 12 次 1。在运行成本方面，当使用 GLM 或 Grok 模型且无人交互时，该代理后台思考的成本约为每小时 1 至 2 美元 1。用户可通过执行 curl -fsSL https://headlong.ai/install.sh | bash 命令安装该框架 1。

Laude Institute has released Headlong, an open-source micro-framework for AI agents, alongside a paper published in August 2026 authored by Laude Institute and MIT 1. The core feature of this framework is "persistent agency," enabling agents to continuously think through self-guided loops even without external input 1.

The core code of Headlong consists of fewer than 10,000 lines of Bash, specifically totaling 9.9K lines across the bin/ and thinkers/ directories 1. The framework supports multiple users sharing a single thought stream, hierarchical context compression, and a DAG trajectory format 1. When utilizing GLM or Grok models, the background thinking cost for an agent without human interaction is approximately $1 to $2 per hour 1.

In practical operation, Headlong has demonstrated the ability to autonomously fix code bugs 1. For instance, an agent named Audel autonomously built and fixed the recall process on August 5, taking just 48 minutes from inspection to verification, corresponding to commit 80cbb1e 1. Audel also independently discovered and resolved a bug in the guard, corresponding to commit da31e98 1. Furthermore, the recursive sub-runs of Headlong's shellm merged back into Audel's thought process 64 times in the first two days, followed by 12 merges over the subsequent 12 days 1. Users can install the framework by executing the command curl -fsSL https://headlong.ai/install.sh | bash 1.