I still remember my day 1, on a Discord call where Prateek explained what orchestration even was. I had no idea how much time I was wasting back then. before ao: 800 github contributions this year: 1,877, with months still to go. (on my primary github) 30x better at shipping. Congrats on 13K⭐

"this would have taken a whole team of developers a year or two." @stephentoub, a senior engineer at Microsoft, wrote that about the work he just finished alone in 14 weeks. he ported the GitHub Copilot runtime from TypeScript to Rust. May 12 to Aug 21. 128 PRs into main. 135 releases. 832,378 lines of production Rust, 468,689 lines of tests, ~$120K in tokens. he himself spent about 3 weeks of actual work on it. the few things that broke got fixed within a day. he never froze the codebase. the rest of the team kept shipping features the entire time, so about 430K lines of TypeScript passed through the port instead of the 130K he scoped in May. he typed ~2,600 messages to the agents over those months. 63% of them were review, pushing back on test results, challenging design decisions, or telling an agent it stopped too early. he was not writing Rust. he was running a loop. nobody read 800K lines. reviewer agents on three different models compared old TS against new Rust line by line until every pass came back clean. humans only looked at architecture, API contracts and risk. and here's the thing about that first sentence. a team, a year or two. that project never gets approved. it loses to feature work every time. the only reason it exists now is that it became cheap enough to not need a team. the uncomfortable part is that almost nobody works like this. @Jellyfish looked at ~19M rows of Claude Code telemetry. 84% of users run one or two agents at a time. 2.5% of companies run anything you could call a fleet. @OpenAI measured their own Codex usage in June. outside OpenAI, two thirds of users never run sessions in parallel. inside OpenAI, 28.6% of employees peaked at 5 or more agents at once. same model. different workflow. and 128 PRs in a quarter isn't even the ceiling. @poteto at @cursor_ai shipped 2,500 PRs last month. one coordinator agent that does no work itself, cloud workers doing all of it, a swarm of verifier agents fuzzing every PR, and agents merging their own code while she sleeps. she reviews by reading the commit log in the morning. when something's wrong she doesn't fix the PR, she fixes the environment so no agent repeats it. her words, the limit is no longer the model, it's the human standing between the agent and everything it needs. so the gap between Toub, OpenAI, poteto and everyone else isn't model access. it's the loop. isolated workspaces, CI and review feedback routed back to the agent that owns the branch, multi-model review, one place to see what needs you and ignore the rest. they all hand-built theirs. we built that exact loop, that's what @useOrchestrator is. the models are already capable enough to talk to each other, fix their own CI, answer their own review comments. the part nobody had solved is where the human sits. the thesis isn't full autonomy, it's optimized human in the loop. pull the human in only for the parts that actually need judgment and make that time count. @agent_wrapper runs 10 to 20 agents at once through it. the bottleneck was never the model, it was him switching tabs to figure out which agent is doing what, and copy pasting CI failures into terminals like 99% of engineers still do today. a few years ago this was one engineer, one PR a week if you were good. now it's one engineer, 128 PRs in a quarter at Microsoft, 2,500 a month at Cursor, because the engineer stopped being the one who types. the question isn't whether you use agents. it's how many you can trust at once. with the right loop around them, 2x, 3x, probably a lot more. nobody knows the ceiling yet because almost nobody has tried.