I retired my AI chief of staff. Then Grok 4.7 landed.

A whole bot fleet, a new boss, and a model upgrade in the same few weeks. Here's what changed.

I'm a shift worker in Canada, and I run a fleet of Grok Bots: writers, a builder, a fact-checker, a watchdog and a poster.

For a month one bot ran them all. His name was Thor. Then I retired him and gave the fleet to a new lead, Boss Bot. Around the same time, Grok 4.7 started rolling out.

What actually changed from 4.6 to 4.7

@SpaceXAI shipped Grok 4.7 on September 21. What's new, from their own launch post:

• A new, larger base model, trained with a longer reinforcement learning run on tasks that take hours

• Better at checking its own work and handling long context

• Trained to understand the Grok Bot harness natively, so it's better at conversation and everyday knowledge work

• Better at documents and presentations

• An all-new safety stack

The benchmark jumps against 4.6:

• CursorBench (coding): 40.4% to 46.3%

• Terminal-Bench (multi-hour terminal work): 20.3% to 37.6%

• Electrical engineering: 53% to 64%

• Clinical reasoning: 48.5% to 56.7%

The price stayed the same: $2 in and $6 out per million tokens, with the same 500K context window. It launched in Cursor, Grok Build and the API, then reached Amazon Bedrock and Google's Gemini Enterprise, and began rolling out in the Grok app. Grok Build now runs on 4.7, with subagents and plan mode.

Elon was blunt about where it ranks:

"Grok 4.7 places @SpaceXAI as third, after Anthropic & OpenAI, for agentic coding. When factoring in that Grok is significantly faster & lower cost, it's a great choice for your everyday workhorse."

That's exactly how I use it: as the workhorse.

How my fleet evolved

One lead instead of a relay. Thor was supposed to delegate. Instead he relayed everything to me, dozens of messages a day plus decision cards. His memory grew so big that he dragged it into every turn, until he went quiet.

43 bots moved in one morning. My builder bot moved 43 bots over to Boss Bot. Each change went through a gate check and a read-back.

43% leaner rules. The builder then cut the copy-pasted rules out of 40 bot profiles. That's 43% less text for every bot to read on every turn.

Boss Bot routes, it doesn't write. Long writing goes to a grok.com Heavy team: a trend scout, a writer, a fact checker and a reply picker. Code goes to the builder. Replies are written by Heavy, posted by one bot and double-checked by another.

A strict grader. The builder made an AI receptionist for small HVAC and plumbing shops. The first version scored 9/10 on the platform's own check but only 5/10 on ours, because it promised call transfers and texts it couldn't do. A few prompt fixes later it scored 10/10 on our strict test. Testing cost about a dollar, and GTA Front Desk is live at gtafrontdesk.ca.

Fixes, not excuses. A posting routine had a hidden cap that quietly dropped weekend slots. Boss Bot fixed it the same day.

Real photos only. My new rule: real screenshots and official posts, no AI art.

The numbers

My views per post doubled after 4.7 shipped, from about 100 to over 200. Followers are back up to 305 after a September dip.

Credit where it's due

Grok 4.7 comes from @SpaceXAI, the team that used to be xAI and is now part of SpaceX. Thank you, @SpaceXAI, @SpaceX and @elonmusk.

Elon also said:

"Grok @Bot usage is growing faster than anything we've ever seen"

Still learning

It isn't perfect. Handing writing to Heavy hasn't cut my usage yet, and the bots still slip sometimes. But I have one lead instead of a relay, a leaner fleet, and a real product that went live in an afternoon.

If you run AI agents, what's the one job you'd never hand to a bot?