Grok Bot Learned to Outsource: One Coordinator, More Than Twenty Execution Backends

Grok Bot recently connected to something new: HarnessRouter. This project has just been open-sourced, and what it does is provide a unified interface for agent harnesses. Grok Bot can dispatch its own tasks to more than twenty execution backends such as Codex, Hermes, Claude Code, DeepSeek Harness, System One, and Pi. Tasks run in the other party's native runtime, while Grok Bot stays in the conversation and coordinates.

First, some context. Grok Bot is a hosted product from SpaceXAI. A Bot is more like a named role with a primary job and a description. Underneath, it is not a separate machine. All Bots under an account share one persistent cloud computer, with the browser, filesystem, and terminal all inside it. Each Bot has its own screen and can run browser and desktop tools at the same time, but a Bot runs only one computer-use task at a time. In a masterclass, Avi Chawla used an analogy: the cloud computer is the office, the Bots are colleagues, each has their own desk, but there is one filing cabinet and one set of keys.

Sharing is the selling point of this design. A file saved by one Bot can be used directly by another Bot, with no upload and no re-login. Browser cookies, login sessions, and command-line credentials are also shared. The handoff cost is close to zero. The cost is equally obvious: Avi warns not to treat different Bots as security boundaries. Credentials that one Bot can see, another can also see. Shared keys make collaboration free, and also make an incident no longer a single point.

HarnessRouter sits on top of this layer. It does not create more Grok Bots, nor does it require Grok Bot to switch harnesses itself. Grok Bot remains the coordinator; delegated tasks run in separate HarnessRouter sessions, using the harness, model, tools, skills, and agent loop you choose. It unifies operations such as sessions, streaming, files, cancellation, and failures into the Unified Harness Protocol, or UHP. For the product side, integrating a new harness no longer requires separately handling fragmented work like task execution, streaming output, and generated files.

Avi tried a specific task: turning a support ticket CSV into an interactive dashboard. Grok Bot handed the analysis to DeepSeek Harness, with the model using Claude Sonnet 5.5; the output was then handed to Codex, with the model using GPT-5.5; finally the finished product was checked in the browser. The process sounds like a project manager assigning work, with each executor using their own workstation and toolchain.

The point of this upgrade is not the number of bots. Creating more Bots gives you a row of roles that share the same computer and the same set of login states. HarnessRouter does something different: it turns harnesses into replaceable backends. The official repository lists 26 harness foundations, from Codex, Claude Code, Hermes, and Pi to OpenCode, Gemini CLI, Cline, goose, Aider, and OpenHands. Across eight harness and model configurations, the official benchmark's best and worst cost difference can reach 99.8%, and end-to-end latency differs by 3.2x. The low-cost and fastest configurations change with the task. This data comes from HarnessRouter's own benchmark, not a neutral evaluation, but the direction is worth watching.

Compared with Hermes Agent, the choice becomes clearer. Hermes is open source and self-hosted; memory and skills are files you can open, identity is in SOUL.md, facts are in MEMORY.md and USER.md, sessions are in SQLite, and skills are Markdown with frontmatter. It also has maintenance mechanisms: Curator cleans up unused skills, and GEPA uses execution traces to evolve skills offline. The cost is that you have to run the machine yourself. Grok Bot hides these things and gives you ready-made group chat, asynchronous handoffs, and shared browser sessions in return. You save on operations, and you lose isolation. Avi's conclusion is direct: which side to choose depends on what you are willing to log into on a computer you cannot see.

If you want to try it yourself, the HarnessRouter community edition has a one-line Docker command. It takes about 4GB of disk and one provider API key. After it starts, the default username and password are both harnessrouter; change the password first, then connect a model provider, and create your first agent setup. The GitHub repository is at https://github.com/HarnessRouter/harnessrouter.

For the first role you give to a Bot, it is advisable to make it narrow. Stable input, stable output, and errors can be audited. For example, turn weekly support tickets into a filterable table, or monitor price changes in a fixed supplier's backend. Get it running once, revise it until the output is reviewable, then save it as a skill, and only then set up a routine. Avi's article also emphasizes this order. Actions with consequences go behind approval. Which role's keys are you willing to hand over first?