✷ AI Open-Source Radar · Aug 17–24, 2026

Agent Harnesses and Constraints Took Five of the Top Ten

Between 2026-08-17 and 2026-08-24, the trend shifted toward the operational scaffolding of agents. Five of the top ten projects focus on agent harnesses or structural constraints, moving the conversation from how to prompt a model to how to build a reliable runtime for it. While most entries target coding agents, MoneyPrinterTurbo sits as a multimodal outlier, automating the entire pipeline from keyword to short-form video.

deepseek-ai/deepseek-harness

DeepSeek Harness is a plugin-based agent harness for building and managing modular AI agents. It solves the problem of rigid frameworks by using a design where every component is treated as a plugin, powered by Cordis for spatiotemporal composability. It now stands at 187,934 stars after adding 42,534 this week. It stood at 83,753 stars on 2026-08-14.

mattpocock/skills

This project provides a set of composable engineering skills for AI coding agents to replace vibe coding with rigorous software fundamentals. By implementing instructions for TDD and domain modeling that can be installed into agents like Claude Code, it attempts to force discipline onto the AI's output. It added 14,911 stars to reach 233,816.

harry0703/MoneyPrinterTurbo

MoneyPrinterTurbo is an AI-driven short-video generator that takes a topic to a finished clip in one click. It handles script writing, material sourcing from Pexels or Pixabay, and voice synthesis. It added 11,105 stars, bringing its total to 115,282.

openai/codex

Codex CLI is a local coding agent that runs in the terminal to provide AI-assisted development. It allows developers to integrate AI directly into their shell environment rather than relying on a web UI. It added 8,915 stars to reach 115,131. It stood at 104,761 stars on 2026-08-08.

stablyai/orca

Orca is a desktop IDE that lets you run many LLM coding agents side-by-side to compare their output and manage the workflow from one app. Unlike the single-agent focus of Codex, Orca uses isolated git worktrees to spawn up to five agents in parallel, allowing you to pick the best result. It added 5,420 stars to reach 51,855. It stood at 31,071 stars on 2026-07-28.

DietrichGebert/ponytail

Ponytail is a framework to stop AI agents from over-engineering code by enforcing a hierarchy of minimal solutions. It implements a decision ladder—checking necessity, reuse, and the standard library before allowing new code—to prevent agents from installing unnecessary dependencies. It added 5,000 stars to reach 108,790. It stood at 89,420 stars on 2026-07-26.

diegosouzapw/OmniRoute

OmniRoute is a unified AI gateway with auto-fallback across 339 providers and token compression to eliminate rate limits. It added 4,866 stars to reach 53,848. It stood at 32,524 stars on 2026-07-28.

earendil-works/pi

Pi is a modular agent harness and self-extensible coding agent CLI. It provides a unified API for multiple LLM providers and a robust runtime for tool calling. It added 4,709 stars to reach 95,906. It stood at 85,501 stars on 2026-08-08.

guillaumemeyer/watermarks-remover

This toolkit strips AI provenance marks and watermarks from text and files. It uses Python scripts for Unicode removal and agent-based paraphrasing to break statistical sampling signals. It added 4,705 stars to reach 17,436. It stood at 2,768 stars on 2026-08-13.

usestrix/strix

Strix is an autonomous AI pentesting tool that validates vulnerabilities with real proofs-of-concept. It added 4,286 stars to reach 57,421. It stood at 44,546 stars on 2026-07-27.

Breaking out

truefoundry/trueforge is an open-source agent harness providing the runtime layer for model calls, sandboxing, and tool execution. It grew by 517.0% to 3,776 stars.

vercel-labs/fx is a minimal and embeddable coding agent harness written in Zig. It grew by 360.5% to 2,238 stars.

llm-as-a-verifier is a general-purpose LLM verification library that scores and tracks AI agent outputs with probabilistic rewards. It grew by 318.3% to 2,673 stars.

s1dashu/ip-as-logo-skill is an agent skill for generating simplified personified mascot logos with strict design constraints. It grew by 311.7% to 3,895 stars.

Leonxlnx/unlazy is an anti-laziness skill for AI agents that enforces task completion through runnable acceptance gates. It grew by 222.7% to 1,665 stars. There is not enough history to judge the shape of its growth.

New to the directory

Developers are moving away from simple prompting toward enforced structural constraints and dedicated runtimes to make agents actually usable in production.