bareagent: Letting Models Decompose Workflows Autonomously

Tight_Heron1730 · reddit · 2026-07-16

This post discusses a key trend: 真正决定 agent 工作流的人是谁, and how the responsibility for task decomposition is evolving from 没人负责 → 人类负责 → 模型负责.

文章中的三种路线

Ralph: No fixed workflow; repeatedly runs a bash loop with the same prompt, using fresh context each iteration to progress.
CodeMachine: Humans pre-define the workflow, and the agent executes accordingly.
RLM: The model treats context as an environment, autonomously decomposing tasks and calling itself recursively, generating its own structure.

作者自己的实现：bareagent

The author implements the RLM (model-led decomposition) approach as a recurse() primitive:

Supports decompose → fan-out → verify → synthesize
Gives the model a spawnchild tool to decide whether to split tasks
Subtasks only receive their local context, returning {result, verdict} or an honest {incomplete}
Clearly flags failures like truncation, refusal, or context overflow without masking them as successes
Uses bareguard to manage budget, rounds, and audit logs

工程信息

零依赖, Apache-2.0
Node >= 18, pure JS + JSDoc
Features 862 个测试
Supports Anthropic / OpenAI / Gemini / Ollama / CLI-pipe
The author also mentions companion projects: litectx (memory), barebrowse (web), and baremobile (Android/iOS)

More from coding & agent