README.md

metadata

title: AGRØN Neural Interface
emoji: 🧠
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
pinned: false
license: mit
short_description: Self-improving multi-agent AI core trained from scratch

AGRØN — Neural Interface

Live deployment of the Agron self-improving multi-agent system.

Agron is an autonomous AI that learns from scratch: it starts with an empty episodic memory and a base LLM trained from scratch on WikiText-2, then improves recursively through observation, memory consolidation, tool evolution and test-gated self-modification.

What runs here

Surface	Description
Command Console	The chat surface. Backed by Agron's own base model, with an optional instruction-tuned voice model when one is available.
Telemetry strip	Live interaction, tool and training counters.
Action log	Timestamped results of every command.
SYSTEM / TOOLS / LEARNING	Diagnostics, the tool registry, and the learning controls (memory training, from-scratch training, dictionary learning, web learning, self-evolution).

Space notes

• Runs on free cpu-basic hardware, so the 6 GB instruction model is disabled by default here (AGRON_USE_VOICE=0). Agron answers with its own from-scratch base model and falls back to a live web-grounded answer when its reply is not usable.
• Auto-push of learned weights is disabled on the Space (AGRON_AUTO_PUSH=0) so the Space's own git checkout is never disturbed.
• Learned weights are published to the model repo teddyoas/Agron.

Configuration

Variable	Default	Purpose
AGRON_VOICE_MODEL	Qwen/Qwen2.5-3B-Instruct	Instruction model used for chat
AGRON_USE_VOICE	1 locally, 0 on Spaces	Run on the native base model only
AGRON_AUTO_PUSH	1 locally, 0 on Spaces	Auto-push learned weights to origin
AGRON_WORK_DIR	/content (/data on Spaces)	Scratch dir for training artifacts
AGRON_PORT / AGRON_HOST	7860 / 0.0.0.0	Bind address
AGRON_SHARE	0	Public gradio.live share link

Run locally

pip install -r requirements.txt
python app.py            # http://localhost:7860