AGI SDK Binary Architecture - Claude Code Pattern
This document describes the AGI driver binary and SDK architecture. The driver is a self-contained agent that captures screenshots, reasons with Claude, and executes actions autonomously. SDKs are thin event wrappers.
┌─────────────────────────────────────────────────────────────────────────┐
TL;DR: Claude Code has a built-in --print --output-format=json mode that bypasses the TUI completely, giving you clean programmatic access to Claude as an async agent runtime.
While building a document editing agent system, I discovered that Claude Code's binary includes a first-class JSON output mode that completely bypasses the terminal UI rendering, making it perfect for programmatic use.
Or: I Accidentally Invented a Way to Make Computers Program Themselves Because I Got Tired of Explaining Things Three Times
Look, I'm going to be honest with you. I built an entire campaign management system—email automation, AI filters, scheduled triggers, the whole nine yards—basically by myself. How? Well, I fired three contractors trying to build it before I realized I could do it myself. By "myself" I mean I got GPT to do it while I ate chips and occasionally typed "looks good, continue."
I haven't actually read the rest of this post yet but I'm pretty sure it's genius. Here you go.
| from abc import ABC, abstractmethod | |
| from typing import ( | |
| Type, | |
| Any, | |
| List, | |
| Dict, | |
| Literal, | |
| Optional, | |
| Set, | |
| Union, |
| GLWT (Good Luck With That) Public License | |
| Copyright (c) Everyone, except Author | |
| Everyone is permitted to copy, distribute, modify, merge, sell, publish, | |
| sublicense or whatever they want with this software but at their OWN RISK. | |
| Preamble | |
| The author has absolutely no clue what the code in this project does. | |
| It might just work or not, there is no third option. |
This visualization was completed for my thesis on measuring the dissapation of transverse sound waves in aqueous media under adiabatic conditions. Jk, its just a cool UV playground
Try it out here!
| async def __call__( | |
| self, | |
| *, | |
| action: Action, | |
| text: str | None = None, | |
| coordinate: tuple[int, int] | None = None, | |
| **kwargs, | |
| ): | |
| if action in ("mouse_move", "left_click_drag"): | |
| if coordinate is None: |
This is a start. It doesn't work atm.
$ python open1.py --epochs 3 --use_mcts --multi_agent --progressive_training --visualize
2024-09-16 07:51:04.977649: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.Objective:
We are building a SaaS application where common features like CRUD operations, authentication, access control, background tasks, real-time communication, and templated entities are required. Below is a general framework you can use for building SaaS applications.
CRUDBase: