Agent Harness

An agent is more than a model call. A useful autonomous agent also needs instructions, tools, state, and a safe workspace.

PyFlue packages these pieces into a Python harness:

1. Load project instructions from AGENTS.md and CLAUDE.md.
2. Load reusable skills from .agents/skills.
3. Create or resume a session.
4. Route the prompt through the selected harness backend.
5. Give the backend controlled access to the sandbox.
6. Persist the result back into the session.

Default Backend

The default backend is DeepAgents:

agent = await init(
    model="openai:gpt-5.5",
    harness="deepagents",
)

The DeepAgents backend provides:

• model selection
• system prompt forwarding
• skill source forwarding
• memory source forwarding
• virtual sandbox adapter
• filesystem permissions
• upload and download file transfer methods
• session thread ids

Pluggable Backends

PyFlue includes a registry so future backends can implement the same public session API:

from pyflue import register_harness

register_harness("custom", CustomBackend)

The current built-in backend names are:

Backend	Status
deepagents	Implemented
openai_agents	Extension point
google_adk	Extension point
pydanticai	Extension point

DeepAgents is the default harness for normal PyFlue projects. The other names are reserved extension points for teams that want to build a custom backend while keeping the same PyFlue project layout and session API.