Quick Agent

Quick Agent is a minimal, local-first agent runner that loads agent definitions from Markdown front matter and executes a small chain of steps with limited context handling. It is intentionally small and explicit: you define the model, tools, and steps in a single Markdown file, and the orchestrator runs those steps in order with a bounded prompt preamble.

Project Goal

Provide a simple, maintainable agent framework that:

• Uses Markdown front matter for agent configuration.
• Runs a deterministic chain of steps (text or structured output).
• Keeps context handling deliberately limited and predictable.
• Supports local tools and simple inter-agent calls.

Install

pip install quick-agent

If you are working from source:

python -m venv .venv
. .venv/bin/activate
pip install -e .

Hello World Example

Create agents/hello.md:

---
name: "Hello Agent"
description: "Minimal example"
model:
  provider: "openai-compatible"
  base_url: "http://localhost:11434/v1"
  api_key_env: "OPENAI_API_KEY"
  model_name: "llama3"
chain:
  - id: hello
    kind: text
    prompt_section: step:hello
output:
  format: json
  file: out/hello.json
---

## step:hello

Say hello to the input.

Then run:

quick-agent --agent hello --input safe/path/to/input.txt

Note: by default, file access is restricted to the safe/ directory (use --safe-dir to change it). Agents can further restrict access with safe_dir in frontmatter (relative to the safe root).

If you omit the entire model: section, the defaults are:

model:
  provider: "openai-compatible"
  base_url: "https://api.openai.com/v1"
  api_key_env: "OPENAI_API_KEY"
  model_name: "gpt-5.2"

Structured Output Example

Create agents/structured.md:

---
name: "Structured Agent"
model:
  provider: "openai-compatible"
  base_url: "http://localhost:11434/v1"
  api_key_env: "OPENAI_API_KEY"
  model_name: "llama3"
schemas:
  Summary: "quick_agent.schemas.outputs:SummaryOutput"
chain:
  - id: summarize
    kind: structured
    prompt_section: step:summarize
    output_schema: Summary
output:
  format: json
  file: out/summary.json
---

## step:summarize

Summarize the input into a short title and 2 bullet points.

Define the schema in src/quick_agent/schemas/outputs.py (example below):

from pydantic import BaseModel


class SummaryOutput(BaseModel):
    title: str
    bullets: list[str]

Then run:

quick-agent --agent structured --input safe/path/to/input.txt

OpenAI API Example

This example uses OpenAI's API via the OpenAI-compatible provider. Set your API key in the environment:

export OPENAI_API_KEY="your-key"

Create agents/openai.md:

---
name: "OpenAI Agent"
chain:
  - id: reply
    kind: text
    prompt_section: step:reply
output:
  format: json
  file: out/openai.json
---

## step:reply

Answer the input in a short paragraph.

Then run:

quick-agent --agent openai --input safe/path/to/input.txt

Python Usage

You can also run agents programmatically:

from pathlib import Path

import anyio

from quick_agent import Orchestrator, QuickAgent
from quick_agent.agent_registry import AgentRegistry
from quick_agent.agent_tools import AgentTools
from quick_agent.directory_permissions import DirectoryPermissions


def main() -> None:
    agent_roots = [Path("agents")]
    tool_roots = [Path("tools")]
    safe_dir = Path("safe")

    registry = AgentRegistry(agent_roots)
    tools = AgentTools(tool_roots)
    permissions = DirectoryPermissions(safe_dir)

    agent = QuickAgent(
        registry=registry,
        tools=tools,
        directory_permissions=permissions,
        agent_id="hello",
        input_data=Path("safe/path/to/input.txt"),
        extra_tools=None,
    )

    result = anyio.run(agent.run)
    print(result)


if __name__ == "__main__":
    main()

How It Works

Agents are stored as Markdown files with YAML front matter and step sections:

• Front matter declares the model, tools, schemas, and chain.
• Body contains ## step:<id> sections referenced by the chain.

The orchestrator loads the agent, builds the tools, and executes each step in order, writing the final output to disk. If the agent front matter omits output.file, the orchestrator returns the final output without writing a file.

Nested Output

When an agent invokes another agent via agent_call or handoff, the nested agent can either write its own output.file or return output inline only. Configure this in the parent agent front matter:

nested_output: inline  # default, no output file for nested calls

Use nested_output: file to allow nested agents to write their configured output files.

Documentation

See the docs in docs/:

• docs/cli.md: Command line usage and options.
• docs/templates.md: Agent template format and examples.
• docs/outputs.md: Output configuration and behavior.
• docs/python.md: Embedding the orchestrator in scripts.
• docs/python.md#inter-agent-calls: Example of one agent calling another.
• src/quick_agent/llms.txt: LLM-oriented project summary and examples.