{"slug": "best-ai-agent-harnesses-and-how-i-use-them", "title": "Best AI Agent Harnesses and How I Use Them", "summary": "A developer's practical review highlights four AI agent harnesses—Pi, OMP, OpenAI's Codex, and Zcode—used daily for coding and automation, with Pi praised as a minimalist open-source option supporting 15+ providers and OMP adding sub-agents and an advisor feature across 40+ providers, while Codex is noted for its Chrome extension and 48-hour autonomous task execution but lacks local model support.", "body_md": "# Best AI Agent Harnesses and How I Use Them\n\nA practical breakdown of the four main AI agent harnesses one developer uses every day for coding, automation, and personal productivity.\n\n## The Four Harnesses\n\n### Pi: The Minimalist Open Source Agent\n\n[Pi](https://pi.dev/?ref=grigio.org) is a minimal agent harness. Small codebase, fully open source, MIT licensed. The best thing about it is that it adapts to your workflows instead of the other way around.\n\nPi ships with powerful defaults but skips features other agents bake in: no sub-agents, no plan mode, no permission popups. You build what you need. Extensions are TypeScript modules that can inject messages, filter history, wire RAG, or implement long-term memory. The agent can even update itself live. Tell it you want to see prices in euros instead of dollars, and it writes an extension and hot-wires it into the footer.\n\nIt supports [15+ providers](https://pi.dev/?ref=grigio.org) including Anthropic, OpenAI, Google, Ollama, and local llama.cpp. The system prompt stays under 1,000 tokens, which makes it fast on local models and cheap on cloud APIs.\n\nThe catch: no built-in sub-agents, no goal tracking, no MCP servers out of the box. You add those yourself or install a package.\n\nGitHub: [earendil-works/pi](https://github.com/earendil-works/pi?ref=grigio.org)\n\n### OMP (Oh My Pi): Pi With Everything Wired In\n\n[OMP](https://github.com/can1357/oh-my-pi?ref=grigio.org) is a fork of Pi that adds everything you're missing. Sub-agents, LSP integration, DAP debugging, hashline edits, browser control, and a model catalog with [40+ providers](https://omp.sh/?ref=grigio.org).\n\nThe killer feature is the advisor. Set an advisor like [Kimi K3](https://github.com/moonshotai/Kimi-K3?ref=grigio.org) and it monitors the main agent's behavior. If the agent goes off track or starts looping in its reasoning, the advisor steps in and corrects it. You can run different models for different roles. Kimi K3 on the API as advisor, GLM 5.3 Flash running locally as the main agent.\n\nOMP also has sub-agents built in. You don't need to set up tmux sessions or run Pi inside another Pi. The harness spawns child processes automatically. Each sub-agent runs with exactly the model and tools it declares.\n\nOne downside: the system prompt is around 30,000 tokens, which slows down prefill on local hardware. On RTX 4600s you get around 2,500 tokens per second. On DGX Sparks it drops to around 1,000.\n\nGitHub: [can1357/oh-my-pi](https://github.com/can1357/oh-my-pi?ref=grigio.org) | Site: [omp.sh](https://omp.sh/?ref=grigio.org)\n\n### Codex: The Desktop App That Does Everything\n\nOpenAI's [Codex](https://openai.com/index/introducing-codex/?ref=grigio.org) is the reference desktop app for agent work. The Chrome extension makes it unbeatable for browser-based tasks. Filing documents, managing email, legal paperwork, research across multiple tabs while you keep working in your own sessions.\n\nCodex can schedule itself to wake up every 30 minutes and continue a goal. Give it a task like \"quantize and prune GLM on the MI300X\" and it runs for 48 hours without interaction. It handles EXL3 quantization, merges models, runs KLD benchmarks, and launches the results. You don't touch it.\n\nThe [Chrome extension](https://chromewebstore.google.com/detail/codex/hehggadaopoacecdllhhajmbjkdcmajg?ref=grigio.org) connects directly to your signed-in browser. It works in isolated tab groups so it doesn't take over your active browsing session. Codex can use plugins first (for services with dedicated APIs), then fall back to Chrome when it needs logged-in browser state, then use its in-app browser for localhost.\n\nThe tradeoff: no local model support. You're locked into OpenAI's ecosystem. If someone built Codex with local models and full provider flexibility, that would be the ideal tool.\n\nChrome Extension: [Codex Chrome Extension](https://chromewebstore.google.com/detail/codex/hehggadaopoacecdllhhajmbjkdcmajg?ref=grigio.org)\n\n### Zcode: The Open Source Alternative to Codex\n\n[ZCode](https://zcode.z.ai/?ref=grigio.org) is Z.ai's desktop app purpose-built for [GLM-5.3](https://zcode.z.ai/en?ref=grigio.org). It's the closest open source alternative to Codex. Goal mode, multi-agent collaboration, browser use, computer use, remote control from your phone via WeChat, Feishu, or Telegram.\n\nThe defining feature is Goal Mode. Instead of issuing prompts, you define a complex objective. The agent breaks it into a plan, executes each step, reviews its own output, and iterates until the goal is complete. Multiple goals can run concurrently.\n\nGLM-5.3 is tuned specifically for ZCode, making agentic coding faster and steadier. But ZCode also supports BYOK for Claude, Gemini, Codex, and OpenAI models. The [GLM Coding Plan](https://zcode.z.ai/en/docs/welcome?ref=grigio.org) starts at $16.20/month.\n\nZCode also has a plugin system since v3.2. Skills, commands, sub-agents, and MCP servers all bundle into one installable unit.\n\nSite: [zcode.z.ai](https://zcode.z.ai/?ref=grigio.org)\n\n## Browser Extensions Compared\n\nAll four harnesses can control your browser, but they work differently.\n\n[SiteGeist](https://github.com/badlogic/sitegeist?ref=grigio.org) is a standalone browser extension built on Pi. It lives in your Chrome/Edge sidebar and works on any website. Bring your own API key from Anthropic, OpenAI, Google, or others. You guide, it executes. No autonomy theater. The codebase is small and clean. You can fork it and have any agent drive it.\n\nThe key difference: SiteGeist is privacy-first. Everything stays on your machine. No account required, no telemetry collected. It's also the only option that works as a pure browser extension without needing a desktop app.\n\nCodex's Chrome extension works in task-specific tab groups. It doesn't take over your active browsing session. ZCode's browser integration goes through the built-in browser or computer use. Pi doesn't ship a browser extension out of the box, but you can install one or build one with the extension system.\n\nIf you're worried about security, use a separate browser for agent work. If you just want something that works with your existing setup, SiteGeist is the best starting point.\n\nSite: [sitegeist.ai](https://sitegeist.ai/?ref=grigio.org)\n\n## Hardware Setup\n\nThe speaker runs a serious local inference stack:\n\n- **4x RTX 4600** for main inference. GLM 5.3 gets around 2,500 tokens per second prefill.\n- **4x DGX Spark** as a cluster. Slower prefill (~1,000 tok/s) but great for running multiple models in parallel.\n- **RTX 3090s** for background tasks: vision, text-to-speech, speech-to-text, embeddings, quantization.\n\nFor a simpler setup with one 3090 and one or two DGX Sparks, the recommendation is to run Qwen 3.8 27B on the 3090 in Q3 quantization for the main agent, and use the Sparks for supporting models like TTS or embedding. Or flip it: put the main agent on the Sparks and use the 3090 for specialized tasks.\n\n## Practical Rules\n\n**CLI tools are best for coding.** If you can see the file tree and read the code, you're more effective. Understand the shape of the repo, not every line. Use Pi or OMP for this. Claude Code works too if you need Fable or Opus.\n\n**Desktop apps are best for long-running goals.** The sidebar lets you jump between active sessions and scale your attention. Use Codex or ZCode for automations, email management, video editing, and tasks that run overnight.\n\n**Browser plugins depend on your threat model.** If you need signed-in state for Gmail, Salesforce, or internal tools, use Codex's Chrome extension. If you want something lighter and privacy-focused, use SiteGeist. If you're on ZCode, use the built-in browser.\n\n**Be careful with the mess.** Agents create files, change commands, write scripts, and generate data. Over time they can degrade your machine if you don't pay attention. Set up sandboxing. Use containers when possible. Monitor what they do.\n\n**You don't need frontier models for everything.** Qwen is enough for most personal assistant tasks: flight searches, hotel bookings, email summaries, document processing. Save the expensive models for coding and complex reasoning.\n\n## The Omar Key Project\n\n[Omar Key](https://github.com/can1357/oh-my-pi?ref=grigio.org) is building an operating system level experience for local AI. The idea is simple:\n\n1. Scan your hardware\n2. Load the optimal model with a pre-prepared recipe and configuration\n3. Open it in any harness: Pi, OMP, Codex, Claude Code, ZCode, OpenCode\n4. Share the server on Tailscale with one click\n\nThe goal is a single-click experience to get local AI running on any machine. It's open source and actively accepting contributions. If you want to shape how operating systems handle AI, this is the place to start.\n\n## Bottom Line\n\nThere's no single perfect harness. The speaker uses Pi and OMP for CLI work and local coding. Codex for long-running automations and browser tasks. ZCode as the open source desktop alternative. SiteGeist for lightweight browser automation.\n\nStart with whatever is open source and least disruptive to your workflow. Add features as you need them. The harness you actually use every day beats the theoretically perfect one you don't.", "url": "https://wpnews.pro/news/best-ai-agent-harnesses-and-how-i-use-them", "canonical_source": "https://grigio.org/best-ai-agent-harnesses-and-how-i-use-them/", "published_at": "2026-09-08 09:51:06+00:00", "updated_at": "2026-09-08 10:02:56.160140+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools"], "entities": ["Pi", "OMP", "OpenAI", "Codex", "Anthropic", "Google", "Ollama", "Kimi K3"], "alternates": {"html": "https://wpnews.pro/news/best-ai-agent-harnesses-and-how-i-use-them", "markdown": "https://wpnews.pro/news/best-ai-agent-harnesses-and-how-i-use-them.md", "text": "https://wpnews.pro/news/best-ai-agent-harnesses-and-how-i-use-them.txt", "jsonld": "https://wpnews.pro/news/best-ai-agent-harnesses-and-how-i-use-them.jsonld"}}