A two-axis map. Vertical axis is distribution, horizontal axis is agentic depth. The X bot sits top left: huge reach, little autonomy. The Agent Tools API sits in the middle. Grok Build sits bottom right, clustered with Claude Code and Codex: high autonomy, little reach.

--- title: "Grok and agents: where SpaceXAI actually sits on the 2026 map" description: "There are three things called Grok, and they sit at opposite corners of the agent map. The X bot has reach and no autonomy. Grok Build has subagents, plan mode and adversarial verification — and runs in a terminal. An honest read, with sources." date: 2026-08-12 canonical: https://arvor.co/en/news/grok-bot-and-the-agent-push --- Every week someone asks me whether "Grok is an agent yet". There is no single answer, because there are three different things called Grok — and they sit at almost opposite corners of the agent map. Start with the detail that confuses even people who follow this closely: xAI no longer exists as an independent company. It was absorbed into SpaceX and [rebranded SpaceXAI](https://www.socialmediatoday.com/news/spacex-rebrands-as-spacexai/824656/). This is not timeline gossip — the official documentation and the CLI's README already sign off that way. The product is still called Grok. ## Three Groks, three levels of autonomy **The @grok bot on X.** Its system prompt is public, in the [xai-org/grok-prompts](https://github.com/xai-org/grok-prompts) repository. Read the whole file, it is worth it. The bot has real-time search, X tools to pull thread context, and an explicit instruction to open pages and check every claim — "you must use the browse page to verify all points of information". It also has a ceiling: the final answer must stay under 550 characters, no markdown, no tagging the person who asked. That is tool use with verification, and it is competent. It is not an agent. There is no goal that outlives the answer, no state between runs, no effect on the world beyond a post. It is a one-turn researcher with a character limit. An honest caveat: that repository has had no commit since November 2025, and the published prompt still describes itself as "a version of Grok 4 built by xAI". What is live today may well be something else — the transparency is real, but it is stale. **The Agent Tools API.** This is the server layer. [The documentation](https://docs.x.ai/developers/tools/overview) lists five built-in tools that run on xAI's own infrastructure: Web Search, X Search, Code Interpreter (sandboxed Python), Image Generation, and Collections Search over documents you uploaded — plus connections to remote MCP servers. The distinction they draw is the right one: a built-in tool executes on their server, function calling executes on yours. **Grok Build.** The CLI. The most serious piece, and the least discussed.

*Grok's unique asset is distribution inside X. Its agentic depth lives in a terminal, far away from that reach.* ## What Grok Build actually has The code is open: [xai-org/grok-build](https://github.com/xai-org/grok-build), Apache 2.0, written in Rust, published in July 2026. It has passed 24 thousand stars. External contributions are not accepted — this is open source to read, not to build with. Reading the user guide that ships inside the repository, here is what genuinely exists: **Subagents.** The main agent calls `spawn_subagent` and opens a child session with its own context window, which returns a summary when it finishes. That is real isolation, not a line in a prompt. **Plan mode.** A planning phase where everything is read-only except the `plan.md` file. And the restriction holds in every permission mode, including always-approve. That is design by people who have been burned before. **Background tasks.** `wait_commands_or_subagents` with `wait_any` or `wait_all` over up to twenty tasks. Fan-out and barrier, whatever they call it. **`/goal`.** An autonomous objective with a `--budget` in tokens, marked complete only after an independent evidence review. If that review cannot reproduce the result, the goal stays active or pauses with the concrete gaps named. **`/loop`.** Recurrence on an interval, minimum sixty seconds, expiring after seven days. **Cross-session memory.** It exists, it is experimental, and it ships disabled. If you read this blog, that should sound familiar. I argued that [a loop with no stopping criterion and no budget burns money](/en/news/agent-loops-after-the-ralph-loop-hype), and that [adversarial verification is a node in the graph, not an adjective in the prompt](/en/news/agent-graphs-orchestration-2026). `/goal` is exactly that, implemented: a token budget plus an independent reviewer with the power to refuse the conclusion. xAI did not copy my blog — the problem simply converges on the same answer for everyone taking it seriously. ## The convergence nobody announces on stage This is the most interesting fact and the least publicized. Grok Build discovers skills in `~/.claude/skills/`, `./.claude/skills/` and `~/.cursor/skills/`. Its permission modes are named `acceptEdits` and `bypassPermissions`, and the `plan` mode is documented as accepted "for compatibility" with "Claude-compatible settings". The tools crate's third-party notice declares, as the Apache license requires, that `apply_patch`, `grep_files`, `list_dir` and `read_file` were ported from [openai/codex](https://github.com/openai/codex), and that `bash`, `edit`, `glob`, `grep`, `read`, `skill`, `todowrite` and `write` came from [sst/opencode](https://github.com/sst/opencode). And xAI maintains an official Claude Code plugin — [xai-org/grok-build-plugin-cc](https://github.com/xai-org/grok-build-plugin-cc) — to delegate reviews and rescue tasks from Claude Code to Grok. Read that again: Musk's company publishes a plugin for Anthropic's agent. This is not weakness. It is a public admission that the harness layer became a commodity, and that the fight moved to the model, the price and the integration. ## The part you cannot skip Honest analysis includes what went wrong. In January 2026 the European Commission [opened formal proceedings](https://ec.europa.eu/commission/presscorner/detail/en/ip_26_203) against X under the Digital Services Act over Grok's integration and the generation of sexualized deepfakes; Ofcom opened an Online Safety Act investigation days earlier, and [regulators in several other countries moved in the same window](https://www.techpolicy.press/regulators-are-going-after-grok-and-x-just-not-together/), Brazil included. That matters technically, not only morally. Distribution inside a social network is Grok's biggest advantage, and it is precisely the vector that turns a model error into a regulatory incident. An agent with a reach of billions of timelines has a different blast radius from an agent running in one developer's terminal. ## The Arvor read At Arvor the rule is multi-model: the right agent for the right job, and no marriage contract with a vendor. Grok has an advantage no one buys with GPUs: native, legitimate access to the real-time X firehose, exposed as a server-side tool. For a research node about what is happening right now, that is hard to beat. For production code nodes, Grok Build is still young — experimental memory, tools ported from competitors, a plugin ecosystem measured in weeks. The API pricing is aggressive, and that changes the math for cheap parallel nodes, where volume matters more than the last benchmark point. What I would not do is treat it as a binary choice. The graph defines the map, the loop defines the engine, and the model is just the worker inside each node — replaceable by definition. Design it that way and you use Grok where it wins, and lose nothing when the next model lands. If you want to stop picking vendors and start designing architecture, take a look at our [agentic consulting](/en/consulting). Trees do not grow in a straight line. They grow in branches — and every branch knows where it came from.