AI Agent HarnessesWhere your agent actually lives

Claude Code runs in a terminal or inside Claude Desktop, reads and writes real files, runs real commands, and schedules recurring work. It defined the skill format that the other harnesses on this page read, which is why it is the reference implementation for most operator kits published today. As of v2.1.302 (October 15, 2026) it ships Claude Mods, the TypeScript based extension system Anthropic announced on October 1, 2026: small functions that hook into Claude Code's events to rewrite prompts before they reach the model, block or retry tool calls, approve or deny permission requests, redact secrets from tool output, and redraw parts of the interface. Mods ship inside plugins, run in the CLI and the desktop app, and are the path the team has been using to ship features like /diff and AGENTS.md support, with more built ins moving to mods over time. v2.1.300 (October 14, 2026) extended the mod system with a prompt.change hook that fires whenever the prompt box's text changes through recall, clear, undo, or a plugin fill, and let a mod rewrite a workflow agent's prompt and model in agent.spawn rather than only refusing; the same release added /advisor on Claude Platform on AWS so an opt in advisor model can steer the agent away from avoidable work, plus OpenTelemetry GenAI semantic convention attributes for trace spans. v2.1.296 (October 7, 2026) earlier added onFailure: "block" for command and HTTP hooks, OSC 7501 working state in supporting terminals, and several session resume fixes. v2.1.283 (September 25, 2026) had added AGENTS.md fallback when no CLAUDE.md is present and shipped plugins through a new developer portal.

OpenAI's agent harness is capable, especially on code heavy work, and the natural pick when the whole team already runs on OpenAI. The Claude Code skill format works there with minor adjustment. Codex CLI 0.159.0 shipped on October 1, 2026 as a major refresh: a new full screen interface, keyboard navigation across older task history, a /voice command for spoken turns, Mermaid and LaTeX rendered directly in the terminal, collapsible diffs and tool output, and a one day follow up in 0.159.1 that made GPT-6.1 Sol the new default model. OpenAI added GPT-6.1 Sol Ultrafast to Codex and ChatGPT Work on October 8, 2026 for Pro $500 and eligible Enterprise and Edu plans, with Enterprise access off by default and workspace owners controlling availability. Earlier in the wave, Codex CLI 0.156.1 (September 23, 2026) had added the first fullscreen TUI, worktree sessions by default, and bundled native voice runtimes.

OpenCode is open source and model agnostic, which is its actual pitch: point it at different models and keep the working setup you already have. Worth it when model flexibility matters more than polish.

Grok Bot arrived in August 2026 as something structurally different: every other harness here runs on your machine, while a Grok Bot gets its own cloud computer. It signs into the tools you use, keeps working while your laptop is shut, and comes back when something needs approval. The October 2026 releases reshaped both its team shape and its price of entry. xAI added Team Bots on September 28, 2026 so a Grok Bot can be given shared files, apps, and expertise and then handed to the rest of the team to work from the same context. The X Subscription on October 12, 2026 folded Grok Bot into a single plan alongside Grok chat, Cursor, and X Premium, with Grok Bot included on the $30 Premium tier for new customers and drawing from a weekly shared pool across Grok, Grok Bot, and Cursor. v0.68.0 added right-click routine controls and clearer approval cards, while v0.68.1 raised the cloud screen to 1920 x 1200, sped up computer actions and screenshots, and added slide deck and formatted email draft workflows. Grok Voice 3.0 shipped on October 14, 2026 as grok-voice-3.0 on the xAI API at the same price as Grok Voice 2.0, so Grok Bot's voice turns now run on the same frontier speech model xAI sells.

OpenClaw is built around always on operation, with recurring work as cron jobs and the agent reachable from the chat apps you already use. The overnight and VPS patterns that operator kits teach map onto it directly. It is the harness this site covers most deeply. On September 29, 2026 the OpenClaw Foundation announced OpenClaw Enterprise (OCE), a vendor neutral, MIT licensed control plane for deploying persistent OpenClaw agents on an organization's own infrastructure, developed in the open with OpenAI, Red Hat and NVIDIA and already piloted inside both Red Hat and OpenAI. OCE adds multi-tenancy, hard security boundaries between trusted and untrusted workloads, LLM based reviews, fine-grained permissions and a standardized agent lifecycle, while leaving the harness, model, and sandbox swappable for third party or internal implementations.

Hermes is the widest open harness in the field: many providers, many chat surfaces, browser and local runtimes, and a local proxy that lets other agent clients borrow its providers. Breadth is the point, and also the cost, so it is a better second harness than a first one.

Pi is a deliberately minimal terminal agent: extensions, skills and prompt templates, and not much else on purpose. It reads skills folders directly and runs headless with one flag, so recurring work goes through cron or Task Scheduler rather than a built in scheduler.

Cline started as the VS Code extension many people met agents through, and now ships a CLI with its own cron for agents that run without a terminal open. Open source, any model provider, and the shortest walk for an editor first workflow.

Alibaba's open source terminal agent is tuned for the Qwen models but takes other providers' keys too. It reads skills, has scheduled tasks and runs headless. The Qwen models are among the cheapest that can do this work, which makes it the budget pick.

DeepSeek's harness runs as dsh and arrived in August 2026 as an open source system where everything is a plugin: models, tools, skills, scheduling, even the interface. It runs a local server with a web interface rather than a chat in the terminal. Newest on this list and moving fastest.

Antigravity is Google's agent first development environment: agents plan, run and verify work across the editor, a terminal and a browser, and report back with artifacts a person can review. Its agy command line runs the same agents headless, which is what makes scheduled routines possible on it. Antigravity 2.17.0 (September 22, 2026) led with 'Plan before you build', and the same week added native Windows support, programmatic turn cancellation, and a nine level MCP safety policy precedence model in the SDK 0.1.18 release on September 21, 2026.

Meta shipped two things in a month: Muse Code, a terminal coding agent, on August 5, 2026, and Muse, a personal agent that lives on a dedicated cloud VM with connectors, skills and crons, on September 8, 2026. Muse Code reads the Claude Code skill format directly and imports existing skills with one command, connects MCP servers, and runs headless with muse exec. Muse itself is the Grok Bot shape: a machine of its own that keeps working after you close the app and comes back when it needs approval.

OpenAI announced dots on September 29, 2026 as always on agents rather than a chat window or a coding terminal. Each dot has its own cloud computer, its own browser and the apps you connect through plugins, over 4,000 of them by OpenAI's count, and you can open its computer at any time to inspect its work. It learns your preferences from feedback, can be called on a voice call, and can be given permission to use your laptop alongside you. Custom Rules let you allow, require approval for or block specific actions, and Activity View shows background work as it happens.