itigges22/ATLAS

Adaptive Test-time Learning and Autonomous Specialization

What it solves

ATLAS enables compact, open-source LLMs to perform complex software engineering tasks locally. It solves the problem of smaller models lacking the inherent reasoning and reliability of frontier models by surrounding the model with a system for planning, candidate generation, verification, and iterative repair, all while running on consumer-grade hardware without relying on external APIs.

How it works

ATLAS implements a multi-layered architecture to enhance model performance:

  • V3 Pipeline: Transforms a single prompt into a verified solution using constraint-driven planning (PlanSearch), diverse candidate sampling (DivSampling), and iterative self-correction via PR-CoT Repair and refinement loops.
  • Geometric Lens: An energy-based scoring system that uses the model's own embeddings and an XGBoost ensemble to predict candidate quality without needing an external oracle.
  • Atlas-Proxy: A Go-based orchestrator that manages tool-call routing, enforces output grammar via GBNF schemas, and applies bias mitigations to steer the model toward structural edits.
  • Sandbox: An isolated execution environment where generated code is compiled, linted, and tested across multiple languages (Python, Rust, Go, C, etc.) before being accepted.
  • Local Inference: Uses llama-server for GPU-accelerated quantized inference, supporting NVIDIA, AMD, Apple Silicon, and Vulkan.

Who it’s for

Developers who want a private, self-hosted coding agent that runs on their own hardware (requiring 16GB+ VRAM) and those who prefer using smaller, open-weight models for software development without per-token costs or data privacy concerns.

Highlights

  • Local-First: Fully self-hosted with no requirement for third-party API keys.
  • Verification Loop: Compiles and tests code in a sandbox to ensure correctness before delivery.
  • Hardware Agnostic: Supports a wide range of GPUs including NVIDIA CUDA, AMD ROCm, Apple Metal, and Vulkan.
  • Native TUI: Includes a Bubbletea-based terminal interface for chatting and monitoring the pipeline in real-time.
  • Model-Agnostic: Allows users to bring their own GGUF models and retrain the Geometric Lens on their own workloads.

Related

  • Project
    esengine/DeepSeek-Reasonix

    An autonomous coding agent that provides a local engine for long-running development tasks with support for checkpoints, sandboxing, and multiple interfaces including a VS Code extension.

  • Dispatch
    AI & Frontier Tech Roundup: Local Models, Agentic Evolution, and the Robotics Race

    Recent developments highlight a shift toward local AI hardware, the rise of self-improving agentic systems, and a strategic push by Chinese labs to challenge frontier model pricing and infrastructure.

  • Project
    1ay1/agentty

    A blazing-fast, C++26-powered terminal coding agent that uses local RAG and a multi-model orchestrator to provide a lightweight, vendor-agnostic alternative to Claude Code.

  • Project
    Q00/ouroboros

    Ouroboros is an open‑source *Agent OS* that adds a specification‑first, replayable workflow to any code‑generating LLM (Claude, Codex, Gemini, Copilot, etc.). A one‑line installer provides the `ooo` CLI; you run `ooo setup` then `ooo interview "…"` to have the tool ask Socratic questions, lock the resulting spec (Seed), generate code, automatically evaluate it, and repeat until the contract is satisfied. All actions are recorded in an immutable ledger, making the process auditable and model‑agnostic. The core kernel lives in this repository, while a terminal UI (`ourocode`) and domain‑specific plugins live in companion repos.

  • Project
    Swival/swival

    A CLI coding agent optimized for reliability with small and local models, featuring broad provider support and advanced context management.