Grok Bot 1.0 isn't new - here's the one upgrade that actually matters | LAB 0042 field notes


Lenny's AI Builders - LAB 0042

the one Grok Bot upgrade that matters, and the paragraph i add to claude.md for long Fable 5.2 jobs.

Grok Bot 1.0 isn't new, but the launch post pulled 46 million views like it was.

so don't worry, you're not missing out.

the one upgrade that's actually decent got something like 40 times fewer views. and it let my bot walk into my Mac.

why it matters

Grok Bot has been out since August. xAI's own page calls 1.0 a major update to its AI teammates. the headline feature, team bots, is for Team and Enterprise plans.

so if you're solo, you got a number, not really an upgrade.

the upgrade that i think really matters is the quiet one. Grok Bot can now join your Tailscale network.

Tailscale is a private network for your own devices. Grok Bot joins it like one more laptop, because it has its own virtual machine. so it can reach things that aren't on the public internet.

what you get

  • the upgrade. what my bot did on my own Mac, with the real numbers.
  • the paragraph. the one i add to claude.md so Fable 5.2 keeps going. you can copy it below.
  • the honest test. i ran it six times. here's what it did and didn't show.

the real example: four doors into my Mac

i put my bot on my private network, where four read-only doors lead into my Mac.

the four doors: my second brain, my agent fleet board, my local models and a health check.

it's a private network only, and every request is logged, so i can check what the bot did.

that morning it counted 42 active agent tasks and listed my four local models in 3 minutes 27 seconds.

and here's the part that wasn't clean:

  • one read out of four timed out twice.
  • i ran it again live at 1:38 and it failed on camera. i left that in.

it's another one of those little one percenter connection improvements. they really do add up over time.

try it

this is the practical, from my Threads post this week. it got 315,000 views. four steps:

  1. switch your model to Fable 5.2.
  2. set the effort level to high.
  3. throw out the effort you tuned on Fable 5.1 or Opus 5.5.
  4. add the paragraph below to your claude.md.

the paragraph is my own trimmed down version of Anthropic's own fix for Claude quitting mid-job.

copy this prompt

add this to your claude.md:

# Finishing work
Do not stop while you still owe work the user asked for. Be persistent. Status notes are fine. If the next step does not need an answer, keep going. For errors, timeouts or failing tools, diagnose first, then retry or use another tool. For a hard blocker (do-not-touch file, safety guardrail), leave it, say what you found, and find another safe way. Stop only when you need user input or the blocker is deliberate. This does not skip confirmation on risky or destructive actions.

check it on a long job

i had my agent test step four. the same job six times on Fable 5.2 at high effort, three runs with the paragraph and three without.

all six finished. so at high effort on this job it made no difference i could measure. i can't give you great proof for it, and that's just me being honest.

Anthropic tested its own longer version at medium effort: about half as many sessions ended with unfinished work, for about 30% more output tokens.

so this is for the super ambitious stuff. check it there:

  • pick a long job, the kind you want to run for days.
  • run it with the paragraph in your claude.md.
  • did it stop while it still owed you work?
  • did it ask before anything risky or destructive? it still should.

limits and dates

  • one bot, one morning. the Tailscale numbers are from my own Mac. it's not a benchmark.
  • one read failed. one read out of four timed out twice, and the live re-run failed on camera.
  • team bots aren't for solo plans. they're for Team and Enterprise plans.
  • my test showed no difference at high effort. six runs, all six finished. Anthropic's number is from medium effort.
  • it costs more tokens. if it's not stopping as much, it's running more. Anthropic saw about 30% more output tokens.
  • the paragraph is my trimmed version, not Anthropic's full text. their full text is in the guide linked below.

the rest of this LAB

  • Gemini 4 Carbon on BuilderBench. i ran Gemini 4 Carbon through my own benchmark and scored its outputs before i saw the result. the score is revealed live at 3:16, so you see it the same second i do.
  • Fable 5.2 takes number one in Agent Arena. Agent Arena ranks agents on real sessions from its community. Fable 5.2 on Max took number one with a 30.4% net improvement, against 19.5% for Opus 5.5 on Max. it's also the dearest: $7.32 a task, against $5.33 for Opus 5.5. Gemini 4 Carbon on high reasoning is at 17% for $1.46. the small print: Arena doesn't define net improvement on the pages i read. Carbon's price is an intro price, and it's reported to double on December 1. my take: i found Fable much more autonomous and proactive than Opus 5.5, but not by much.
  • Gemini Decision 1. Google launched a model that doesn't write. you give it a question and a list of 2 to 255 answers, and it picks one. up to a thousand questions a request, $0.10 per million input tokens, and the output is free. an answer comes back in about a fifth of a second, and it's free to try in the API. where a builder uses it: sorting tickets, routing leads, yes or no checks inside an agent. anything where the answer comes off a list. it's not the first of its kind. Jev had early access on September 15, OpenAI's Decisions API came on September 29, Microsoft on October 9 and Google on November 11. my take: hype for now.
  • the GPT-6.2 rumour. a leaker on X said the GPT-6.2 family launches Thursday, US time, aimed at Opus 5.5 and Fable 5.2 at once. OpenAI hadn't announced anything when i filmed. it was still a rumour. i'll run any new OpenAI model on BuilderBench. my guess: it won't be an Opus or a Fable killer, but let's call it a noteworthy rival.
  • Nvidia's board and Jensen. when i filmed, the Wall Street Journal had reported that Nvidia's board was preparing to remove Jensen Huang as CEO. it was a report, not a result. update since filming: on November 12 Nvidia announced that Jensen Huang resigned as CEO, and CFO Colette Kress is interim CEO. why should you care as a builder? a lot of the AI in this episode runs on this one company's chips.

sources

next step

watch LAB 0042 on YouTube - the practical starts at 10:47.

want the next one? get the next field notes. optional - these notes stay free to read.

Lenny’s AI Builders

i’m Lennox. i sift through AI Twitter and share one thing i tried with ChatGPT. for people making products, content and useful systems. five emails a week, with a weekly option.

Read more from Lenny’s AI Builders

Lenny's AI Builders - LAB 0045 one Claude Code setting, 36 to 38% fewer tokens in every pair of my test, the catch on cost, and a prompt that gives you 10 versions. if you build with Claude Code, one new setting can make it use fewer tokens. i tested this myself. every pair of runs used 36 to 38% fewer tokens. but there is a catch. my first run cost 83% more at api prices. the other two pairs came out 15% and 17% cheaper. why it matters the setting turns on Code Mode. Claude writes code that...

Lenny's AI Builders - LAB 0041 the conductor paragraph, the four steps to run it, and the honest numbers from my 40-hour run. i ran my viral Fable 5.2 conductor for 40 hours on my real app. one conductor, 73 worker threads, 25 pull requests merged. and my weekly Fable limit was gone in 13 hours. so here's what it actually built, and the one paragraph that runs yours. why it matters this is the new meta. you use Fable 5.2 as the conductor and Opus 5.5 as the worker threads. Fable 5.2 is twice...

howdy so i let Claude reread six months of my agent logs. (yes, i called it dreaming) that became the centre of this LAB Roundup. my agent's memory went from 7 to 9 Claude proposed 125 edits to the notes. then i ran the 10-question quiz i wrote before the test. the score moved from 7 to 9 out of 10. but i still had to check the work: 54 edits held up, 58 only partly, 2 did not, and 11 had no log day to check. watch LAB 0039 · listen the video skill found 7 of 14 cuts i gave it my best edit...