From scattered elements to structure.
Plutonus is live.
Products, client work, experiments, and the systems behind them — all in one place.
https://t.co/7XzXmtjk5z
@ClemensScharti@typesafeai@antigravity Yeah, this is a much cleaner boundary. 👍
User intent can remove the annoying confirmations for things they actually asked for, without becoming a blanket pass. Catastrophic ops still stopping for confirmation feels right.
@mukundzha6 I checked the write path after seeing the “every write shows as a diff before it touches disk” line. Looks like new files skip the preview — write_file checkpoints them and writes directly. Am I reading that right?
@YevMur@nick_realm_01 What happens when the agent has to edit a file that should keep the real secret locally? Do you replace it with a placeholder before the request leaves the machine, then restore it before the edit is applied?
@thomasjankowski@stark0xbt The verification cost has a nasty tail. I lost ~50 minutes to one green agent PR because the agent had rewritten the fixture until the failing test passed. The run looked cheap; figuring out whether the change was actually valid was the expensive part.
@jaketselby@omarsar0 Most cross-runtime tooling seems to stop at “we rendered the right config.” This one explicitly refuses to call that a runtime proof. Good distinction.
The handoff boundary here is really well chosen. Shared objectives, decisions and artifacts can cross runtimes, but native memory, approvals and verification claims don’t magically become portable state.
Especially since the receiving runtime has to re-run the gate rather than inherit “verified” from the previous one.
@pauliusztin_ The two-queue split makes sense. I’m curious about the safe-injection boundary though — do you only allow steering between tool/model turns, or can the harness decide mid-turn that a message can still change the current task?
@thefrikidude That autonomy boundary makes sense to me. Routing, risk, and recovery can stay cheap, but execution only crosses over for known verification commands.
Are those commands from a fixed allowlist, or can JEV construct or select them dynamically?
@DanKornas I like the frozen-candidate constraint.
You can spin up parallel agents pretty easily. Keeping one version in place long enough to actually review it is a different problem.
@bbelderbos Yeah, that’s the part I’m still testing.
My current line is whether I can still push back on the code, change direction, and catch the agent when it’s wrong.
If I can’t, I no longer own it.
@ClemensScharti@typesafeai@antigravity That user-intent case is especially interesting.
“Unsafe because the agent decided to do it” and “unsafe, but exactly what the user asked for” probably shouldn’t be treated the same way.
@pidotdev This is a big part of why Pi clicked for me.
Very little harness, very little ceremony, and still enough surface for the model to do real work.
@Marko_Poly I ran into the same problem with external state itself: if you keep everything, it just becomes another transcript.
What helped was separating what’s true now, what happens next, and what we observed — then deleting anything superseded.