autourgos-preiteration

Pre-iteration middleware for Autourgos agents. Run any sync or async callback — and inject files like screenshots — before every agent iteration. Built-in image compression keeps vision LLM costs low.

from autourgos_preiteration import PreIterationMiddleware
from autourgos_agent import Agent

middleware = PreIterationMiddleware(callback=capture, files="/tmp/screen.png", image_quality="low")
agent = Agent(llm=my_llm, middleware=[middleware])
result = agent.invoke("Open the browser and search for Python 3.13 release notes")

Features

• Fresh context every iteration — screenshots, live data, health pings, whatever your callback produces
• SEQUENTIAL/PARALLEL combinators for running multiple callbacks
• Sync and async callbacks, both work inside invoke() and ainvoke()
• Dynamic file injection — pass a callable for files to generate a fresh path per iteration
• Built-in image compression (via optional Pillow) — from a flat ~85 tokens ("low") to no resize ("auto")
• Zero required dependencies; works with any Autourgos agent

Table of Contents

• Why Use This?
• Install
• Quick Start
• Run Multiple Callbacks
• Async Callbacks
• Dynamic File Injection
• Image Quality
• Parameters
• License

Why Use This?

Some agent tasks need fresh context every iteration:

• Computer-use agents — take a screenshot before each step so the LLM sees the current screen state
• Live data agents — refresh a price feed, sensor reading, or API response before each reasoning step
• Monitoring agents — ping a health endpoint or log iteration metrics before the LLM call

PreIterationMiddleware handles all of this cleanly, with zero boilerplate inside your agent.

Install

pip install autourgos-preiteration

For image resize/compression support (Pillow):

pip install 'autourgos-preiteration[images]'

Quick Start

from autourgos_preiteration import PreIterationMiddleware
from autourgos_agent import Agent

SCREENSHOT = "/tmp/screen.png"

def capture(iteration: int) -> None:
    take_screenshot(SCREENSHOT)  # your screenshot function

middleware = PreIterationMiddleware(
    callback=capture,
    files=SCREENSHOT,
    image_quality="low",  # ~85 tokens flat — great for computer-use agents
)

agent = Agent(llm=my_llm, middleware=[middleware])
result = agent.invoke("Open the browser and search for Python 3.13 release notes")
print(result)

With Agent(verbose=True), this middleware also narrates its own actions into the agent's trace:

[PreIteration] Injected 1 file(s) before iteration 2.

Run Multiple Callbacks

SEQUENTIAL — one after another

from autourgos_preiteration import PreIterationMiddleware, SEQUENTIAL

def capture_screen(iteration: int) -> None:
    take_screenshot("/tmp/screen.png")

def log_step(iteration: int) -> None:
    print(f"Starting iteration {iteration}")

middleware = PreIterationMiddleware(callback=SEQUENTIAL[capture_screen, log_step])

PARALLEL — all at the same time

from autourgos_preiteration import PreIterationMiddleware, PARALLEL

def refresh_cache(iteration: int) -> None:
    cache.clear()

def ping_health(iteration: int) -> None:
    requests.get("https://api.example.com/health")

middleware = PreIterationMiddleware(callback=PARALLEL[capture_screen, refresh_cache, ping_health])

Async Callbacks

import asyncio
from autourgos_preiteration import PreIterationMiddleware

async def async_capture(iteration: int) -> None:
    await asyncio.sleep(0)  # non-blocking
    take_screenshot("/tmp/screen.png")

middleware = PreIterationMiddleware(callback=async_capture)

Works inside both agent.invoke() (sync) and agent.ainvoke() (async).

Dynamic File Injection

Pass a callable for files to generate paths per iteration:

def get_screenshot_path(iteration: int) -> str:
    path = f"/tmp/screen_{iteration}.png"
    take_screenshot(path)
    return path

middleware = PreIterationMiddleware(files=get_screenshot_path, image_quality="medium")

Image Quality

Resize requires Pillow: pip install 'autourgos-preiteration[images]'. Without Pillow, the image_detail hint is still applied, but no resize/recompress happens.

Parameters

License

Apache License 2.0, Copyright (c) 2026 Jitin Kumar Sengar