Kelven Severo

Just building.

[Kelven Severo @KelvenSeveroF]

@wewals the glow-up is they learned when to stop. never ask was the pitch, taste is knowing which actions still need a human

[Kelven Severo @KelvenSeveroF]

@Dennis_Rye that's the worst place a remembered yes can land. i'd deny-write that settings file and treat any secret that landed in the repo as already leaked

[Kelven Severo @KelvenSeveroF]

@KI_Vater @NousResearch freezing a good run is the right cost cut. i'd still hash the inputs though, otherwise a stale freeze keeps shipping yesterday's answer

[Kelven Severo @KelvenSeveroF]

@stretchcloud the tell is it couldn't say where the extra account came from but the block was already live. a destructive action needs a dry-run diff against the source list, not a self-grade after the fact

[Kelven Severo @KelvenSeveroF]

@waldekm i'd be careful with retries in the same session. after two misses the model often just rereads the failures. a clean session with a tighter prompt is the fairer test

[Kelven Severo @KelvenSeveroF]

@brendenvonhouse that's the quiet compatibility layer. an agent that has to screenshot a pricing table will misread a footnote and pick the wrong tier

[Kelven Severo @KelvenSeveroF]

@yog_codes yeah async total is the move. show page 1 immediately and let the count land when it lands. users notice a spinner on first paint way more than a missing 1 of N

[Kelven Severo @KelvenSeveroF]

@ganesh2412 yes. red for some other reason just proves git revert works, not that the test can catch that bug. pinning the original failing fixture is what makes the replay real

[Kelven Severo @KelvenSeveroF]

@ohheyyang yeah if it can edit its own history you're reviewing a story it wrote. the log has to live somewhere the agent can't touch

[Kelven Severo @KelvenSeveroF]

Self-edit without a locked core is just self-corrupt

Exo freezes history and policy activation in a protected layer while the rest of the harness rewrites itself as code.

Cost drops 35-61% on AppWorld and SREGym. The fence is the product.

[Alex Krentsel @AlexKrentsel]

Agent harnesses are usually hand-tuned, or optimized offline against benchmarks. What if an agent could improve its own harness *while doing the work*?

Introducing Exo, a harness architecture for runtime self-improvement 🧵

(paper: https://t.co/0PZQms3Jo0)

[Kelven Severo @KelvenSeveroF]

Green tests are the alibi for skipping the diff

Addy Osmani almost pushed PRs to close competitor gaps after reading the research and skimming the code. He called approving a 600-line green diff code ratification.

If green is why you approved it, you didn't review it.

[Addy Osmani @addyosmani]

https://t.co/KD5BxbuZ5Q

[Kelven Severo @KelvenSeveroF]

@lateascent i think plain users notice fast. the cursor jumping while you type is hard to miss, and that's when most people would just turn it off

[Kelven Severo @KelvenSeveroF]

Task success is the wrong score for computer use agents

Cua's bench only counts a task done in the background while you keep working. Best on macOS fully passed 58%. Codex computer use 33%.

If it needs your mouse, you didn't hand off the work. You just stopped yours.

[Cua @trycua]

1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv.

Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.

[Kelven Severo @KelvenSeveroF]

@ohheyyang yeah the split matters more than the score. a stuck agent costs you a minute, a confident wrong one costs you an afternoon

[Kelven Severo @KelvenSeveroF]

@markfenner @qoder_ai_ide the leftover worktrees and docker stacks bit is underrated. multiple agents leave a mess after the pr lands

[Kelven Severo @KelvenSeveroF]

@buildAgoat a date picker inside the thread beats describing a calendar to a text box every time. the card is the ui

[Kelven Severo @KelvenSeveroF]

@brendenvonhouse 180 links in llms.txt aimed at the agent is the right instinct. most docs still assume a human will click around

[Kelven Severo @KelvenSeveroF]

@VibeCoderOfek schema edits nuking the cache for the rest of the session is a silent bill. fresh thread after big schema moves

[Kelven Severo @KelvenSeveroF]

@iam_ihoudjedje mit keeps the binary. it doesn't keep your setup when the cloud kills the worktrees and secrets. export path first

[Kelven Severo @KelvenSeveroF]

@surlymochan claimed done without a replayable proof is just a green checkbox. independent evidence should be the default gate

[Kelven Severo @KelvenSeveroF]

@ishaansehgal @omnaraai recovery mid-command is the real gap. wait vs retry vs collect is where hosted harnesses quietly lie to you

[Kelven Severo @KelvenSeveroF]

@witcheer 142 combined is cool. for coding the number i care about is what one session gets when three others are also generating