Build a Coding Agent

This guide builds a small coding-agent workflow with a Markdown skill and a typed result.

Create the Project

pyflue init code-agent
cd code-agent

Edit the Skill

Open .agents/skills/triage.md and use:

---
name: triage
description: Inspect a request and decide whether a safe fix was applied.
output_schema:
  type: object
  properties:
    fix_applied:
      type: boolean
    summary:
      type: string
  required: [fix_applied, summary]
---

# Role

You are a careful Python coding agent.

## Instructions

Inspect the request, make only safe changes, and summarize the result.

Write the Agent Script

import asyncio

from pydantic import BaseModel
from pyflue import init


class TriageResult(BaseModel):
    fix_applied: bool
    summary: str


async def main():
    agent = await init(
        model="openai:gpt-5.5",
        harness="deepagents",
        allow_write=True,
        allow_shell=True,
    )

    session = await agent.session("coding-agent")
    result = await session.skill(
        "triage",
        args={"request": "Review the repository and suggest a safe fix."},
        result=TriageResult,
    )

    if result.fix_applied:
        await session.shell("git status --short")

    print(result.summary)


asyncio.run(main())

Notes

• Shell execution requires allow_shell=True.
• File writes require allow_write=True.
• Use command grants for production workflows, for example allowed_commands=["git", "pytest"].
• Keep commits behind explicit approval or application logic.