xAI's Grok Bot team published its 72-hour live-build playbooks, 69 agent roles and 40 documented failures as an MIT repo

Three engineers from the xAI Grok Bot team, Roshan Sadanani on product, Lauren Tan on engineering and Matt Palmer on developer experience, published the internal materials they used to build and launch a product from an empty repository in 72 hours, live on stream, using their own agent platform. The repository at unicodef1wn/grokbot-field-notes has reached 411 stars and 58 forks in three days, under an MIT licence that lets any team copy the files into its own project.

What is in the repository

The main file is an AGENTS.md that turns a coding agent into a verification loop rather than an open-ended writer. Alongside it are 69 agent roles, one file each, and 9 role playbooks for engineering, product, founders, sales engineering, sales, SDR, support, post-sales and marketing. A separate failure log records 40 antipatterns the team met during the live stream, each paired with the rule the team wrote to stop it happening again. The pull-request template included in the repository turns the verification loop into a form with four required fields: what was reproduced, what proof is attached, what scope the change touches and what needs a human to sign off.

What the team says worked

The published rules keep the agent inside a loop of small, provable steps. The team writes that every change has to reproduce the problem before it proposes a fix, and has to attach the evidence in the pull-request body. Two PDFs come with the repository, a 24-page main guide and a 14-page marketing-focused guide, and both are checked in so a reader can read them without running the agent.

What the team says did not work

The 40 antipatterns are labelled as things that broke during the live stream, which makes the failures the most useful part of the repository for another team, since they were documented in the moment rather than reconstructed afterwards. The team also states one strict rule: keep a human gate on migrations, deploys, money and permissions, no matter how well the loop has been working. The team is clear that this rule stayed in place through the whole 72 hours, and that it should stay in place for anyone copying the AGENTS.md into a repo where those four categories exist.

A team building with a coding agent today can take the AGENTS.md, the pull-request template and the 40-item antipattern list and add them to its own repository this afternoon. One limit matters. The numbers, roles and rules describe a product the xAI team wrote in three days, so the same rules may need adjusting for a codebase that already carries production traffic. The failure log is where a team should read first, because a rule is easier to keep once the failure that produced it is on the page next to it.

Source

• unicodef1wn: unicodef1wn/grokbot-field-notes on GitHub