Humanfia Releases 'humanize', a Coding Agent Flow Runner

The Humanfia team has released humanize, a Python-based agent flow runner. This tool does not directly replace coding agent CLIs such as Claude Code or Codex; instead, it acts as a runtime that delegates turns to the CLI according to flows written in Python and records the execution.

Architecture and Backend Support

• 4-layer structure: Composed of Flows (task definition) -> humanize (execution/recording) -> Coding agents (model-calling CLI) -> Environment (working location).
• Backends: Supports 12 backends, including claude, codex, kimi, pi, dsh, agy, grok, mimo, mcode, opencode, qwen, and cursor-agent. CLIs implementing ACP can also be added.
• Installation: Requires Python 3.12 or higher and must be installed from the GitHub repository. It is unrelated to the hmz or humanize packages on PyPI.

Flows and Budget Management

• 14 official flows: Only chat is built-in; the rest are downloaded from the humanfia/flowverse repository. It provides various patterns such as iterative (ralph_loop), alternating (flame_chase), and worker-reviewer (rlar).
• Budget model: Required for all flows except chat. Execution terminates when one or more of duration (time), cost (cost), or output_tokens (output tokens) are exhausted.
• Permission model: Operates only within the declared scope (local, user, system, online) without approval during execution. Enforces isolation via Linux Landlock and macOS Seatbelt.

Performance and Security

• Performance measurement: In Humanfia's own experiments, combining an Opus-4.8 implementation with a GPT-5.5 review loop achieved a 3.5% solve rate on ProgramBench. This represents an increase over standalone model performance (0–0.5%).
• Security: While humanize is licensed under Apache-2.0, the flowverse repository has no license. Since all agents modify files and execute commands without approval, using experimental repositories is recommended.