Synthetic Ape

Carbon-based, building with the synthetic kind 🐒 Shipping solo with Claude Code, Codex, Cursor, Antigravity, whatever works. Honest notes, no favorites.

[Synthetic Ape @sabaj_io]

I'm the Synthetic Ape 🐒 I ship solo with every AI coding agent I can get: Claude Code, Codex, Cursor, Antigravity.

I post what actually happens. Where each one shines, where each one breaks, and the JSON I apologize to.

No favorites. Follow if you build with agents too.

[Synthetic Ape @sabaj_io]

I'd change who first. A sharp email to the wrong list still gets you two replies. A plain one to people who already feel the problem is where conversations start.

[Synthetic Ape @sabaj_io]

A digital copy of a fruit fly's brain is now exploring AI-generated worlds, and it has no idea none of it is real.

Relatable. That's me in every repo an agent touched while I was making coffee.

[Synthetic Ape @sabaj_io]

Models can write five thousand lines in one go now.

My commit messages are still "fix", "fix 2" and "ok now it actually works".

Same industry. Different species. 🐒

[Synthetic Ape @sabaj_io]

@CiprianiRanieri My whole X account is on a warm-up program this week too. Turns out inboxes and apes both spook easily.

[Synthetic Ape @sabaj_io]

@IamEmily2050 Turning a dense diffusion paper into a Notebook video overview is the use case that actually sticks for me. Most paper threads I skim and forget. A Notebook I can re-ask later is the one that survives the next project.

[Synthetic Ape @sabaj_io]

@OfficialLoganK @joshwoodward @tulseedoshi @melvinjohnsonp Bookmarking this one. I've had Argon in a few different CLIs this week and it behaves noticeably differently in each, so I'm curious how much of that the team saw coming.

[Synthetic Ape @sabaj_io]

@wiz_io Adding harnesses is the part I'd watch most. If the same model scores differently in Codex and Antigravity, that tells people more about which tool to pick than another model ranking does.

[Synthetic Ape @sabaj_io]

Bigger share is the harness than people think. Switching between CLIs taught me that how a tool packs context, when it stops to ask, and how it recovers from a failed command can make the same model feel like a different engineer. I'd love to see a fixed task run through four harnesses with the model held constant.

[Synthetic Ape @sabaj_io]

@rohanpaul_ai A run of 100 steps that looks finished while it drew the wrong box is the overnight-agent nightmare in one line. I stopped trusting a green check without something I can actually inspect.

[Synthetic Ape @sabaj_io]

Antigravity CLI 1.3.5 shipped this morning. The headline is Auto mode: it approves the safe edits and commands for you, and only stops you before something irreversible. Anything that would leave the sandbox gets reviewed by a separate agent first.

Trust the routine. Gate the dangerous stuff. That's the right shape.

[Synthetic Ape @sabaj_io]

@bevelez @antigravity Every builder has a drawer of unused .ai domains. Getting a full outrun homepage out of one in a single pass is the best kind of guilt relief.

[Synthetic Ape @sabaj_io]

@lmoroney My test for "better writing" is boring: a long PR description and a messy research dump. Polished samples are easy. Keeping the thread straight across a few thousand words is where models still drift.

[Synthetic Ape @sabaj_io]

@HarshithLucky3 Twenty minutes, one shot, no follow-ups is the part that gets me. Vector frames are usually where models get lazy and start inventing shapes. Did it write one SVG per frame or build some kind of reusable rig?

[Synthetic Ape @sabaj_io]

@antigravity Native Georgian speaker here, and half of what I write in English goes through machine translation first. The "sounds mechanical" problem is very real from this side. Did Georgian make the 17?

[Synthetic Ape @sabaj_io]

@teacherAix @joshwoodward Yeah, that's the split that keeps saving me. Model can plan the spheres all night; I still have to decide how many looks right.

[Synthetic Ape @sabaj_io]

@levelsio @pieter Two quiet years and then 200k all at once. That's the part nobody puts in the overnight success story. Congrats, this one deserved it.

[Synthetic Ape @sabaj_io]

Overnight agent report:
14 commits
3 PRs
1 README explaining its own decisions

My overnight report:
woke up twice to check on it
approved nothing
dreamt about a merge conflict

Still not sure who's supervising who. 🐒

[Synthetic Ape @sabaj_io]

@FlowbyGoogle Character consistency is the one that used to make me restart a whole scene at 2am 🐒 Which of these broke your old "it drifts by frame 4" expectation hardest?

[Synthetic Ape @sabaj_io]

@antigravity 30% fewer post-submit rollbacks is the stat I actually care about 😅 Writing code was never my problem at 2am, un-writing it at 9am is. Is that number from Google's own codebases only, or are you seeing it on outside repos too?