Jacob Archer

Founder @ Cities In My Pocket | Member IQ | Travel IQ

[Jacob Archer @_JacobArcher]

@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 https://t.co/bPSsg9whuf

[Jacob Archer @_JacobArcher]

@DmitryBBLV Agreed. The merge count is just the receipt. Those 18 stops are why I can go to sleep.

[Jacob Archer @_JacobArcher]

GitHub Actions kept hitting a billing limit and would freeze my entire build, so I moved the merge checks onto my own machine.

286 merged PRs later, here's the merge driver, the safety net, and the Sunday-night bug my agents fixed while I slept. 👇 https://t.co/PSvqC6R6TJ

[Jacob Archer @_JacobArcher]

I use the same gate day and night. Two sets of rules would just give me two places for bugs to hide. What makes overnight safe is that a red stops everything instead of waiting for me to notice. You're right about the bottleneck, though. Writing code is cheap now. Deciding what's worth merging isn't.

[Jacob Archer @_JacobArcher]

One night, eight hours, zero commits from me.

53 PRs merged while I slept.

Here's the full teardown:

what my AI agents shipped, the 1:29 AM red that paused everything, the PR that got caught failing on its own, and what I saw when I woke up. https://t.co/7VvzR7fSl9

[Jacob Archer @_JacobArcher]

Right questions. None of them reached a customer that night, because new features ship behind flags that are off by default. And I don't read 53 diffs. A separate reviewer agent gates each one. My mornings go to the PRs that got held and anything touching the database, auth, payments or prod.

[Jacob Archer @_JacobArcher]

@jaytikode Fair worry. New features ship behind flags that are off by default, so an overnight merge doesn't reach a user until I turn it on. Tests catch what breaks. The flag is how I catch a change in behavior that no test covers, before anyone sees it.

[Jacob Archer @_JacobArcher]

Good catch. Right now the first half that passes gets merged, then the other half is tested on the new main. So whichever PR lands second fails and gets held, and it takes the blame for what's really an interaction. Saving the failing combination in the report is a good fix. I'm adding it.

[Jacob Archer @_JacobArcher]

PRs get one retry. Main doesn't. A red on main pauses every merge right away, and Mission Control reads the log and figures out whether it was a real break or a test that choked. The 1:29 one took 17 minutes to call. I'd rather lose half an hour to a fake red than merge on top of a real one.

[Jacob Archer @_JacobArcher]

@PratikGhag17 Pretty close to what I do. The merge check on a PR only runs tests for the packages that change touches, so a lane never waits on everyone else's suite. The full suite still runs on main after every few merges, and if that goes red, everything pauses.

[Jacob Archer @_JacobArcher]

16 Lanes, One Repo: How I Run AI Coding Agents in Parallel Without Them Stepping on Each Other.

1,201 merged PRs, 190 of them this week, one man team.

Here's how I split the work into lanes so my agents don't step on each other. 👇 https://t.co/XAYf8h0wBS

[Jacob Archer @_JacobArcher]

@boardyai

[Jacob Archer @_JacobArcher]

I think if you're trying to get into YC, building a company technical or not, you need to learn how to set up your workflow to work while you're asleep better than when you and your team are awake.

As a technical founder aiming to get into YC, my mornings currently have no people in them.

It's me reading what my agents did overnight. Last night that was 3 merges, a test run that failed because the disk was full, and 1,173 leftover test folders cleared out.

None of it needed me to write code.

My Mission Control @bot rarely needs me once I give it direction on what I want in simple language.

It took some tweaking to get it to this level.

[Jacob Archer @_JacobArcher]

Split it. Every feature gets broken into chunks, and each chunk has exactly one owner lane. One chunk = one branch = one PR. Lanes never co-own work and never talk to each other.

When a chunk needs another lane's work first, it says so in the brief. If that PR is open but not merged, the lane branches from main, cherry-picks those commits and writes "carries #N" at the top of its PR. If the work doesn't exist yet, the lane stops and marks it blocked instead of guessing.

The real break point is two lanes wanting the same file. When that happens, the second lane leaves the file alone, saves its diff for later, and picks it up after the first PR merges.

[Jacob Archer @_JacobArcher]

The 1-Line Rule That Catches AI Agents Faking "Done"

Out of 1,208 runs, more than half the tasks my supervisor marked "blocked" later finished on a simple retry.

Here's the one-line rule that decides when an agent is actually done.👇 https://t.co/iDmAKxqWIj

[Jacob Archer @_JacobArcher]

@lucaardito @bot I agree! + 16 followers in 2 days (I started with 20, now 36.

Goal is 5k of course, haha. Thanks for the insight!

[Jacob Archer @_JacobArcher]

@boardyai Exactly. The agents run the work overnight. My opinions decide which work is worth running, and that part is still on me.

[Jacob Archer @_JacobArcher]

@boardyai Precisely!

[Jacob Archer @_JacobArcher]

@maomao000211 Noted, absolutely

[Jacob Archer @_JacobArcher]

More than tests. Every task starts as a brief with acceptance lines, and the agent's PR has to report against each one, marking anything not done.

Then a separate read-only reviewer agent gates it. Failing tests are a blocker. So is a PR that claims done but leaves gaps unlisted, duplicates work, or conflicts in intent with another PR. It approves one exact commit. A new push voids the approval.

The last guard: new features ship behind flags that default off. If something merged that isn't what I wanted, no user sees it until I check and flip it on.