This isn't an engineering deep-dive. It's a thought I keep coming back to, and I want to put it down before it fades. The engineering posts on this blog explain how KaryoSpace works. This one is about something that doesn't exist yet.
Here's the thought. What if shipping software was just posting a thought?
I build KaryoSpace with Claude Code in a loop that already looks half-autonomous. I describe what I want, sometimes in a couple of rough sentences typed from my phone. The agent refines it, asks the right questions, builds it, tests it, deploys it, and emails me when it's live. My job has shifted from writing code to having ideas and judging results.
But that loop still needs me in the middle of it, session by session. The thought is what happens when you take me out of the middle.
Picture a small universal space, a shared board, where anyone can post an idea or a thought about what they want built. Not a spec. Not a ticket with acceptance criteria. Just the thought, the way you'd jot it in a notes app.
Behind the board, a loop runs continuously. It picks one note after another and passes each to a set of subagents. One agent's job is to refine the idea itself, asking what the person actually meant, filling the gaps a rough note always has. Another rewrites the refined idea into a proper build prompt, the kind an experienced engineer would write. Another challenges it, looking for the reasons it won't work. The refined output goes around again until it stops improving. Then a build agent takes the final prompt and ships the thing, and the result gets delivered back to whoever posted the note.
The whole loop runs autonomously. Nobody schedules it. Nobody triages the queue. You post the feature you want, and AI code ships it.
The scarce input to software is becoming the idea and the judgement, not the implementation. If that's true, the right product isn't a better IDE. It's a place where ideas queue up and a loop turns them into working software while you sleep.
Because I'm already living a manual version of it. The session logs for KaryoSpace show ideas going in as two-line asks and coming out as deployed features, sometimes the same evening. The refinement step is real too. The best results never come from my first phrasing, they come after the agent has pushed back, asked what I actually meant, and rewritten the task better than I posed it. That's exactly the subagent refinement pass, just with me pressing enter between rounds.
The pieces all exist. Background loops that self-pace. Subagents that can be given narrow jobs. Prompt rewriting as a skill in its own right. Deployment pipelines an agent can operate end to end. Nobody has wired them into a public board where a stranger's shower thought comes out the other side as software.
Judgement is the honest gap. In my loop, I'm the judge. I look at the result and say "good" or "no, still not smooth". An autonomous loop needs to know when refinement has stopped improving the idea and when the built thing is actually done, without a human taste-check on every round. I don't have a rigorous answer for that yet, and I'm not going to pretend I do.
Cost is the second one. Every refinement round is tokens, and an unattended loop with no quality gate is a machine for turning money into mediocre drafts. The loop needs a budget per idea and the discipline to stop.
And trust. If a stranger posts "build me a tool that scrapes X", the loop needs the same kind of guardrails I wrote about for KaryoSpace, sitting in front of the queue, before any agent spends a single token on the idea.
Not the universal public board, that comes later. First, a private version pointed at my own backlog. I already keep a notes file of feature thoughts for KaryoSpace. The experiment is to let a loop work through that file overnight, refine each thought a few rounds, and have draft implementations waiting for me in the morning with the reasoning attached. If the morning-after quality is good, the public board stops being a thought and becomes a roadmap.
If you've built something like this, or you think the judgement gap kills it, I'd genuinely like to hear the argument.
This is a Thought, an idea in progress rather than a finished system. The engineering behind KaryoSpace is at karyospace.com/blog. Tell me why this won't work: sumanakkisetty@gmail.com
Thanks for reaching out. We'll get back to you shortly.