Flue is a TypeScript-based framework for building autonomous AI agents using a programmable “agent harness” architecture. It is designed to give developers full control over how agents plan, execute tasks, manage context, and interact with tools, rather than relying on fixed SDK abstractions or prebuilt agent behaviors.

The core concept behind Flue is “Agent = Model + Harness.” Instead of treating an AI agent as just a language model with prompts, Flue provides a structured runtime layer (the harness) that includes sessions, memory, reusable skills, and tool execution capabilities. This allows agents to perform complex workflows such as reading files, running shell commands, spawning sub-agents, and coordinating multi-step tasks.

Flue enables developers to define agents as programmable functions in TypeScript. Within these functions, agents can call structured “skills” (reusable workflows with typed outputs), maintain persistent sessions, and execute commands inside secure sandbox environments. These sandboxes can be local, virtual, or connected to external container providers, giving agents a safe environment to interact with code, files, and systems.

A key design goal is portability. Agents built with Flue can be deployed as HTTP servers, run via a CLI, or executed inside environments like Node.js, Cloudflare Workers, GitHub Actions, and CI/CD pipelines. This “write once, deploy anywhere” approach makes it suitable for both local automation and production infrastructure.

Flue emphasizes developer ownership of the full agent stack, including the harness logic, sandbox configuration, and security boundaries. Sensitive credentials such as API keys can be managed with fine-grained control, ensuring that agents can execute tasks without directly exposing secrets to the model runtime.

The framework supports building a wide range of agents, from simple chat-based tools to complex systems like coding agents, data analysts, customer support bots, and automated issue triage pipelines. Its architecture aligns with emerging “agent engineering” practices, where reliability, structure, and execution control are prioritized over prompt-only approaches.

Key features include:

• Programmable agent harness with sessions, memory, and skills
• TypeScript-first development model for defining agents
• Secure sandbox environments (local, virtual, or container-based)
• Structured skill execution with typed inputs and outputs
• CLI and HTTP server deployment options
• Integration with CI/CD systems (GitHub Actions, GitLab, etc.)
• Fine-grained control over environment variables and secrets
• Support for multi-step workflows and sub-agent coordination
• “Write once, deploy anywhere” architecture

Common use cases include:

• Building autonomous coding and debugging agents
• Automating issue triage and DevOps workflows
• Creating AI-powered data analysis pipelines
• Developing customer support agents with knowledge bases
• Running agents in CI/CD pipelines and cloud environments
• Prototyping and deploying custom agent-driven applications

Flue is developed by the Astro team and positioned as an “agent harness framework,” focusing on giving developers full control over agent behavior, execution environments, and deployment infrastructure.