Thirty Days of Grok Heavy Build

Running Grok 4.5 Build under SuperGrok Heavy, inside a Claude-native harness I started in December 2025 · July 8 – August 8, 2026

I spent the first ~30 days of a SuperGrok Heavy subscription running Grok 4.5 Build — the agentic CLI — inside a harness I built first for Claude Opus, then integrated Codex, and later other model routes for volume versus judgment.

This is not a marketing feature list. It is what the on-disk session store shows when every Heavy chat history in that window is inventoried and pushed through a redaction-aware mine.

The month, in one table

Headline: Heavy Build was not a chat toy in my harness. It was a fleet co-processor — thousands of headless workers, hundreds of interactive sessions, and a habit of treating “done” as something I re-check on disk.

Names

A foreign harness, on purpose

Most demos put a model in its own living room. I did the opposite: a system designed around Claude and Codex, with Grok Heavy as a high-capacity, tool-capable family and multi-account routing. Cheap volume went to detached workers; judgment and single-writer artifacts stayed with me.

If Build only works inside a pure xAI demo loop, it is a product. If it thrives inside a foreign, fail-closed operator stack, it is infrastructure.

Here, it proved to be infrastructure.

Scale is a catalog

I did not ask a model how many sessions I had. A script walked the Heavy session store, classified working directories, and wrote a catalog.

Most sessions were not long chats with me. The median was a bead: a detached, prompt-filed worker writing into an isolated output directory, then harvested by me in a controlling session that does not trust a “complete” flag without re-reading the files. That is the product story — Build as bacterial labor, not only as a pair programmer.

Interactive sessions were the steering surface: automations, meters, design, games and visual critics, capacity routing, adversarial verifiers. One peak day (2026-07-28) was almost entirely that kind of work — a studio day and a control-plane day sharing one account envelope.

Multi-track month

Capacity retrospectives fail when they collapse parallel programs into one story. This month ran simultaneous tracks: fleet orchestration; content and education; games and visual QA; systems truth (usage meters, multi-model accounts, host load, routing).

Keyword co-occurrence across takeaways (not exclusive labels) put verification and orchestration first, with content, capacity, games, and infrastructure close behind. A single session can hit several themes.

What Build was good at here

Long-horizon work with tools. Multi-hour arcs: research → filesystem → reports → notes. Not just single-turn Q&A.

Headless volume without losing the plot. Workers for scouts, competitive profiles, fix passes, reviews — with isolation, registration, and a re-check of the files before I accept them.

Adversarial self-check. Separate Build sessions tasked to refute “objective complete.” Optimism is cheap; refutation is culture.

Capacity as a portfolio. Heavy is a metered weekly envelope you aim, not a bottomless key. Account routing and burn honesty mattered as much as prompts.

Scars worth shipping

1. “Complete” is a claim. Disk re-check is the gate.
2. Prose is not state. Counts live in catalogs; status prose is a pointer.
3. Parallel authors fracture voice. Coupled narrative stays sequential.
4. Packaging ≠ publish. A human stamp is separate from a finished draft.
5. One account blocked is not a fleet halt. Re-route; re-observe live meters.
6. Host surge is real. Wide fan-out under heavy load manufactures flakiness.
7. Redaction before fame. Full-session mines touch private work; public layers only consume what is tagged safe.

How this review was made

Honesty about method is part of the story:

1. Deterministic catalog of every Heavy chat history in the window
2. Deterministic digests (title, asks, theme hits, sensitivity heuristic) for all of them
3. Fifty size-balanced shards → cluster → chronology and capability integrate
4. One sequential write of this piece

Why digests before generative judgment: feeding multi-gigabyte raw JSONL into dozens of concurrent models under a loaded host is a reliability failure dressed as ambition. The pyramid still exists; the base is evidence-compressed.

Roughly half the takeaways stayed out of the public chain under redaction heuristics (including known false positives). Better blocklists can refine later without re-mining raw text into drafts.

What I would tell a new Grok Build user

1. Put Build in a real harness — output contracts, hard checks before you call work done, notebooks.
2. Prefer detached workers for volume, interactive for judgment.
3. Inventory sessions before storytelling.
4. Treat capacity as a portfolio, not a single chat.
5. Ship practices in the same motion as demos.
6. Never auto-publish from agent enthusiasm.

Closing

Thirty days in, Grok 4.5 Build under Heavy is not a sidebar experiment in my operator OS. It is a primary pair of hands for fleets, a critic for its own optimism, and a co-author of durable practice.

The piece is the month: thousands of sessions, one redaction-aware pyramid, and a public story that refuses to invent what the disk does not show.