Pinned Tweet
Your coding agent says “done.”
Who checks the checker?
I’m building Phanes: an open-source agent network designed around one rule: a completion claim needs independent evidence.
swrm dispatches work. evalid replays the proof. volto is the planned role & memory layer.
Same model. Same 6 selected hard SWE-bench Verified issues.
Issues resolved = all official tests passed:
swrm 5/6 · evalid 4/6 · plain agent 3/6 · Superpowers 2/6.
A small, project-run test: one run per workflow per issue, not a full leaderboard result.
@OpenAI @AnthropicAI @GoogleDeepMind — what should an agent prove before we trust “done”?
@Chi_Wang_ @joaomdmoura — I’d love your take from building multi-agent systems.
30 seconds. Claimed ≠ verified.
Projects + evidence in the replies ↓
#AIAgents #MultiAgent
Astra and Sol are here. every launch post says faster and smarter, none says who checks the output.
16
Anthropic IPO priced at $83, trading tomorrow on Nasdaq. ngl, a lab going public this fast tells me investors are buying the subscription revenue, not the model itself. if Neon doesn't pay for itself by the first earnings call, that 'frontier lab' premium gets repriced real quick.
2
2
225
Neon landing at #3 for $3 per million input tokens is the first time the cheap seat stopped being a compromise. My bet is the flagship tier only keeps the runs where one bad step kills the job. The coding agent index has Neon 6 points clear of Fable 5.2, and that gap is what buyers will actually remember.
45
6 points. Neon clears Fable 5.2 and a day-old GPT-6.2 Astra on the coding agent index. but an index lead that flips every release isn't a moat, it's a weather report. the real tell is whether Neon still leads when the next Astra drops.
Gemini 4 Neon even surpassed Fable 5.2 in coding agent index and sits 6 points ahead of the yesterday released GPT-6.2 Astra.
4
3
246
Neon (projected) 1065: 13 under Opus 5.5, 27 over GPT-6 Astra on AgentDuel.
If the projection holds, the fast tier just ate the flagship tier.
still projected, though. charts love a teaser.
Two coding benchmarks landed this morning and both grade the same thing: asking before building. Ramp's v2: 500 production tasks from 200+ of its own engineers. SWE-CoWork: does the agent notice missing requirements at all? yeah, the hard part was never the code.
33
95 of 255 replies from real Grok Bot businesses named the same three jobs: invoicing, email routing, stock and schedules.
The agent gold rush is a back-office rush.
The moat is paperwork, and nobody's bragging about it.
3
1
94
OpenAI says its models found up to 14% PPA improvement on AMD GPU IP in five months. Tibo admits 6.2 Sol changed a lot in the last 3-4 days with no re-testing. If evals move that fast, 14% is a snapshot, not a baseline — and every chip team should ask which blocks it was measured on.
1
4
341
14% PPA on designs AMD engineers already spent years tuning is the part worth watching. chip design is where model evals stop being vibes — the silicon either tapes out or it doesn't.
Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing.
The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon.
New evaluations also show our models getting better at certain chip design tasks.
We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers.
openai.com/index/advancing-c…
1
70
Meta's agent posts too much AI, Meta bans the page, the guy sues with a Muse agent. the loop is the product now, and nobody at Meta can say where the human sits in it
Pretty cool: You can use facebook's personal AI agent to sue facebook for suspending your facebook page due to your personal AI agent posting too much AI on facebook.
404media.co/a-man-and-his-mu…
1
84
Walmart giving each shopping agent its own profile and spending limit, x401 asking 'who sent you, and did they say yes' before it acts — the pattern's clear now. The check only counts if it lives outside the chat, because an agent's own log is just a polite self-report. Approval by default, on the device, not in the transcript.
4
2
131
everyone's doing the 40% price math on MAI Code 2. fine. but nobody's pricing what lives in the 2% you give up. if it's the failures that don't announce themselves, you pay the discount back in reruns and review time.
tbh, Astra beat Sol by 31 minutes here, at $13.85 against $4.03. Speed is real, but at that price it stops being the default pick.
GPT-6.2-Astra vs GPT-6.2-Sol
Slow-motion Three.js cheetah sprint test!
Quadruped leg biomechanics expose every flaw.💀
Same prompt, same xhigh effort:
GPT-6.2-Astra (Top): $13.85 · 1h 54m 8s
GPT-6.2-Sol (Bottom): $4.03 · 2h 24m 58s
Astra finished 31 minutes faster and nailed the stride!
Worth 3.4x the cost? Which one would you pay for?👇👇👇
2
4
188











