Flue: Open-Source Agent Harness Framework

Current

Flue: Open-Source Agent Harness Framework

Flue is an open-source agent harness framework developed by an Astro co-founder, enabling headless AI agent creation via minimal TypeScript and Markdown-based logic definitions for tasks such as code triage and translation.

Signal

NEW: Astro Co-Founder Unveils Flue, Open-Source Framework for AI Agents · twitter · 2026-05-02 Flue is introduced as an open-source agent harness framework developed by an Astro co-founder. It enables developers to construct headless AI agents using minimal TypeScript code and Markdown files for logic definition. The framework targets tasks such as code triage and translation, positioning itself as a tool for creating lightweight, file-driven agent behaviors without complex configuration overhead.

Context

Flue emerges within the agent harness pattern, a structural layer that abstracts agent execution mechanics, tool binding, and state management. By utilizing Markdown for logic definition, Flue adopts a declarative approach that prioritizes human-readable agent specifications. The framework's foundation in TypeScript suggests a focus on type safety and developer ergonomics, while the involvement of an Astro co-founder indicates potential alignment with modern web development workflows. The emphasis on headless execution makes Flue suitable for automated environments where UI interaction is unnecessary.

Relevance

Flue contributes to the agent infrastructure layer by offering a harness implementation that reduces friction in agent development through file-based logic. The use of Markdown for behavior definition supports version control, code review, and static analysis of agent rules, addressing common challenges in agent reliability. This pattern reinforces the filesystem-native agent state circuit, where agent configuration and logic are treated as persistent, versioned artifacts rather than ephemeral code constructs.

Current State

Flue is available as an open-source framework for creating headless AI agents. It supports task execution in domains such as code triage and translation, demonstrating applicability to development automation workflows. The framework combines minimal TypeScript code with Markdown-based logic files, allowing developers to define agent behavior through a hybrid approach that balances type safety with declarative simplicity.

Open Questions

How does Flue manage dynamic tool discovery and binding when logic is defined in static Markdown files? Does the framework support multi-agent orchestration, or is it limited to single-agent workflows? What is the runtime mechanism for parsing Markdown logic, and how does it integrate with standard tool protocols such as MCP? How does the framework handle state persistence and recovery across agent invocations?

Connections

Flue operates as a specialized agent harness that emphasizes file-centric logic definition. Its Markdown-based approach resonates with the filesystem-native agent state circuit, where agent behavior is encoded in persistent structures. The framework's headless design and TypeScript foundation align with terminal-native agentic workflows, providing a structured method for agent integration into automated pipelines and CI/CD systems.

Connections

  • Filesystem-Native Agent State Infrastructure - Aligns with file-based logic definition pattern; uses Markdown files to encode agent behavior, reinforcing the circuit's emphasis on persistent, versioned file structures for agent configuration. (Circuit · en)
  • GitAgent Protocol - Shares the principle of defining agent behavior via open, readable formats, supporting cross-runtime interpretability of agent logic through declarative specifications. (Current · en)
  • Missing connection:

Related entries

External references

Score

Score derives from linkage, recency, and abstract depth; at-risk merely suggests erosion and does not indicate retirement.

Mediation note

Tooling: OpenRouter / qwen/qwen3.6-flash

Use: drafted entry from external signal, assessed linkage against existing knowledge base

Human role: review, edit, and approve before publication

Limits: signal content may be incomplete; verify primary sources before publishing