Another week, another terminal coding agent. xAI shipped Grok Build on May 14 — parallel workers, a plan mode, a CLI. My first reaction was to skip it; the AI coding space ships a “new” agent roughly monthly, and most of them are a thin wrapper and a waitlist. My second reaction, after actually reading the feature list, was that the feature list is the story — just not in the way xAI’s launch post intends.

The shape has converged

Line up what shipped in the last six weeks:

• Grok Build: eight agents in parallel, a 2M-token context window, a plan mode you approve, comment on, or rewrite, and changes delivered as reviewable diffs.
• Claude Code Agent View (May 11): background sessions, one unified list of every running and waiting session, plan-then-execute.
• Cursor Composer 2: parallel agents running inside the editor against your open files.
• Zed 1.0: parallel agents plus the Agent Client Protocol, an attempt to make the agent layer pluggable.

Four serious tools from four well-funded teams, and the interaction model is now identical: you describe a task, the tool produces a plan, you approve or edit that plan, several agents execute in parallel, and you review the result as diffs. Eighteen months ago the live differentiator was whether a tool could reliably edit more than one file in a single turn. Now multi-file, multi-agent, plan-gated execution is the floor. The headline features of a 2026 launch are indistinguishable from the headline features of every other 2026 launch, which means the launch post has stopped being useful information.

That convergence is itself the signal. When four independent teams land on the same shape, the shape is no longer a bet — it is the settled answer to “what does an agentic coding tool look like.” Categories do this right before the competition moves from “who has the features” to “who executes them best,” and that second contest is much harder to win with a press release.

When the shape converges, the moat moves

If every tool has the same workflow, the workflow is not what you are choosing between. The decision retreats to three places, and it is worth being specific about each:

1. Model quality on real tasks. Not the benchmark, not the scripted launch demo — your codebase, your gnarly four-hour refactor. This is the only axis where a raw newcomer can win on day one, because it is the only one that does not require an installed base.
2. Ecosystem binding. Copilot is wired into GitHub: pull requests, Actions, review in the diff view. Cursor is wired into the editor you already live in. Claude Code is wired into the Anthropic surface and the skills ecosystem accreting around it. These bindings took years.
3. Price predictability. Flat versus metered, and how high the ceiling climbs before a finance conversation starts. This is quietly becoming the sharpest axis of the three.

Grok Build competes almost entirely on the first point, plus a bet that 2M tokens of context behaves like real working memory rather than a number on a slide. It has no editor, no Git host, no install base. That is a hard corner to launch from, because moats two and three are precisely the ones a brand-new entrant cannot build quickly — they are the ones that take years and a userbase to accumulate.

The $300 anchor

SuperGrok Heavy is about $300/month, discounted to $99 for the first six months. Set the discount aside; introductory pricing is a customer-acquisition tactic, not a signal. The list price is the signal. xAI is pricing access to a coding agent at the level of a high-end professional tool — a seat of a serious commercial suite, not a $10 add-on to something you already pay for.

That is not an outlier. Anthropic’s Max plan and GitHub’s move to usage-based Copilot billing point the same direction: the market is actively testing how high the ceiling goes for an agent that genuinely does multi-hour autonomous work. The implicit claim behind a $300 price tag is “this replaces enough of an engineer’s time to be worth a real fraction of an engineer’s tooling budget.” Maybe it does. But the 2.3-tool developer I described last week — the one already running Cursor, Claude Code, and Copilot in parallel because each holds a slot the others cannot — is now being asked whether one of those slots is worth $300 on its own, on top of what they already pay. That math is far harder than “$10 well spent on autocomplete.”

Does a fourth player change anything?

Mostly no, and it is worth being honest about why. Developers do not adopt a CLI agent because it benchmarks well in a launch post. They adopt it because it is already wired into something they use, or because it is unambiguously better at the work they actually do — often enough that switching pays for the friction. On day one Grok Build is neither wired in nor demonstrably better. It is a strong-sounding spec sheet from a vendor with no coding-tool install base and no editor to fall back on.

There is exactly one dimension where it could matter. If Grok 4.3’s quality on long, multi-step agentic tasks is materially ahead of the field, and the 2M context holds up under load instead of degrading into the usual mid-context amnesia, Grok Build becomes the “hard task” tool in a multi-tool stack — the slot Claude Code holds for most serious practitioners today. That is the only slot a no-ecosystem entrant can take, because it is the one slot decided purely by model quality. It is also, for exactly that reason, the slot most exposed to whoever ships the next better model next month. Winning it is not a moat; it is a lease. Whether Opus 4.7 already signed a longer lease on that slot is the question I take up in the stack reshuffle.

The spec sheet also hides a cost it will never advertise. Every agent you add to the stack is another mental model to maintain, another set of failure modes to learn, and another place your own judgment can quietly erode while the agent does the thinking. The supervision problem does not get easier when you bolt on a fourth tool — it gets one tool harder, and the marginal tool has to be very good to be worth that.

A crowded race full of nearly identical cars is decided by the engine and the pit crew, not the bodywork. Grok Build has, on paper, a loud engine and, for now, no pit crew and no home track. That is survivable — Cursor started with no install base either, and won the editor slot by being decisively better at it before anyone else took the slot seriously. But it means the launch spec, however well it reads, is the opening line of the argument, not the close of it. For the release details and pricing, see xAI ships Grok Build.