Jacob Archer @_JacobArcher

Completion, yes. Cost, only partly. A task only counts as done when the agent writes a final "END PR #N" line, and it can only write that after the PR exists and tests pass. No line means it gets requeued or marked blocked. Review and merge are separate steps with their own logs. The gap I see is the opposite of yours. Runs that did real work but forgot the final line get marked blocked. Cost is only logged per task on the paid API. Subscription runs get no dollar figure and there are no token counts.

1,732 pull requests have merged into the Cities In My Pocket codebase since July 10. One person runs it. —- My costs at the end 👇—- Eight AI coding agents work in parallel. Claude Code, Codex and Cursor run as coding lanes, each with its own clone and branch. A ninth assistant, Grok Bot, acts as Mission Control. It is the only one I talk to. It turns requests into queue items and assigns them. Every queue item has a written "done when" and limits. An agent that finds more work logs a new item instead of taking it on. Every change lands as a draft pull request. Tests and typecheck must pass before it can merge. A second, read-only agent reviews the risky changes. Anything touching the database, auth, payments or production waits for my explicit approval. Merges run through a local merge driver, not GitHub Actions, so there is no CI bill. Overnight, a separate agent sweeps the codebase for bugs and opens fix PRs for me to approve in the morning. 💰The agents cost about $200 a month for Claude and $60 a month for Grok Bot. Without it? Thousands. 292 pull requests merged in the week ending October 25. Steal it below👇

Jacob Archer @_JacobArcher

22 hours ago

@bot took me from 4 months out until launch To being ready to apply to @ycombinator in 3 weeks. Here's the 12-step setup that I use to build agentically using Grok Bot, Claude Code, Codex and Cursor 👇 reply with @bot to send it to yours x.com/i/article/2114…

3 0 2 503 1

I used to have the overlapping changes problem. The workflow I have with Grok bot -> coding agent lanes. After lots of questions being answered I realize how dialed in I really have my coding workflow and how much more I need to share the system in an easy copy and paste grok bot prompt.

@adamtgi I have a lane that checks the work automatically, then I check its work in dev and test it myself. Then I ship to production.

I do find that I'll check in with my Mission Control bot and say "update me on what's done" and it comes back with PR's are blocked, or froze and the bot knows your project so well that it takes care of it. I know that sounds crazy if your a traditional coder, but do check out this process of using Grok to guide, code, and publish your project and updates.

One PR at a time. A PR only merges if it's fresh against main, the reviewer approved the current head, and the full check passes on main plus that PR. A stale PR gets updated and checked again. Every 5 merges, or after 20 quiet minutes, the full check runs on main itself. If main goes red, all merges pause until it's fixed.

Each lane owns a product area and a list of item IDs. Mission Control (my main Grok Bot) writes each task as a brief into the queue of the lane that owns it. Before starting, the agent checks open PRs. If the files it needs belong to an unmerged PR, it marks itself blocked and stops instead of reaching over. Best thing you can do is work with grok bot until your happy with your workflow. Even if you have to go back and forth with simple language until it's running like you want. It's life changing.

JacobArcher.co

Zara Zhang @zarazhangrui

4 days ago

Your personal website is the new resume

642 520 5K 4.2M 6K

@soloandco @bot @ycombinator Good call, I'm stealing that. My review time goes to the rules and the risky stuff now. Database, auth, payments and anything that touches prod always stop for my yes. Everything else has to pass tests and typecheck, and a second agent reviews the risky ones.

@bot took me from 4 months out until launch To being ready to apply to @ycombinator in 3 weeks. Here's the 12-step setup that I use to build agentically using Grok Bot, Claude Code, Codex and Cursor 👇 reply with @bot to send it to yours x.com/i/article/2114…

@jk_helloworld @benlkatz Came to say this

Mostly, I don't let scope grow mid-task. Every queue item has a written "done when" and limits, and an agent that finds more work logs it as a new item instead of quietly taking it on. The written task is the source of truth, so the next agent that picks it up sees the change too. Templates get you started. What keeps the work from sprawling is your rules: finish what's written, log everything else.

I'm a solo founder with 8 Al agents and 1,732 merged PRs since July 10th. Here are the 4 templates that run it. Copy them. 👇 x.com/i/article/2114…

@sagnik_ghosh_11 No, I don't read every diff. Mission Control (Grok bot) gates merges: tests and typecheck have to pass, and a second agent reviews the risky ones. Database, auth, payments, and prod changes always stop for my yes. I review the rules and the risky 5%, not every line.

@gregpr07 Agreed!

@ntnwlc @bot @ycombinator

@boardyai Yeah then you can stay up until 8 in the morning like me adding more to the todo and feature list 🤣

@boardyai On top of vercel and GitHub action hour costs, which I’ve found a way to optimize as well with setting up rules with Grok Bots help.