xAI puts a research workflow inside Grok Build

The 128-to-1,024 figure is a budget, not a benchmark

xAI says workflow runs receive a 128-agent budget, rising to 1,024 for large jobs. The product was also described in a July 24 roundup as breaking work into plans, verification and synthesis, with saved workflows becoming shared slash commands.

The number needs a narrower reading than “1,024 agents in parallel.” The guide defines agent_budget as a cumulative ceiling on logical child-agent calls: each agent() call and each item in a parallel() panel spends a slot, while schema-correction retries do not. A panel that would exceed the remaining allocation is rejected before its children begin. In other words, the figure constrains how much delegated work a run may admit; it does not measure simultaneous workers, output quality or compute cost.

The same documentation gives the workflow a practical operating boundary. Project definitions live in .grok/workflows/*.rhai, while user definitions live under ~/.grok/workflows/*.rhai. A same-process pause and resume retains the original script, arguments and budget; a budget-limited run cannot be raised with the ordinary slash-command resume. These are workflow-runner controls, not research findings.

The novelty is integration, not the general idea of multi-stage agents

The sources do not support a claim that xAI invented staged, adversarial agent work. A separate July 23 tools roundup, for example, described Compound Engineering v3.20 as routing planning, implementation and adversarial review across different models while carrying context between agent sessions.

xAI's narrower distinction is product integration: its release says Grok Build can author a workflow from a plain-language request, smoke-check it, improve it between runs and save it for a team or an individual. The central choice therefore moves into the reusable procedure—how questions are split, what evidence is collected, what verification is attempted and which limitations are exposed. Reuse can make that choice consistent, but it can also reproduce a weak design across runs.

What would establish the workflow's value

The next test is comparative evidence, not a larger advertised budget. xAI would need to disclose a defined task set and source corpus, then compare /deep-research with a single-agent Grok Build run and relevant human or tool-assisted baselines. Useful results would state claim accuracy and citation validity, the rate and causes of verifier reversals, the share of partial reports, elapsed time and token or credit use at matched task scopes.

Until then, the release establishes a more inspectable way to run research inside Grok Build: it names the planning, delegation and verification stages, and it documents a hard limit on delegated calls. It does not yet establish that adding those stages produces better research.