Walrus

real setups, live dashboards. dm open.

[Walrus @Walrus_systems]

It is becoming f*cking independent

Andrej Karpathy’s Claude Code setup just showed Duet running a full field with the laptop shut.

Operator asleep. Claude keeps working. Nothing gets sent.

The field in the video:
- Market field live across crypto and weather books (sol perp, eth perp, doge spot, rain london, heat phoenix, wind north sea, frost calgary)
- Equity from seed $50 to held $5,273
- Laptop shut · Claude keeps working
- allow: read · sort · own notes
- ask: send · invite · move
- deny: pay · delete · secrets
- “nothing sent without your yes”
- Field live · asks before it acts
- Positions open, floors watched, kill switch ready, human approvals at zero

Most agents either stop when you close the lid or keep trading until something breaks. This one runs the book and waits for yes.

Anthropic’s version is still a research preview. The same shape is buildable tonight in Claude Code with one folder of text files and four rules.

[Walrus @Walrus_systems]

It is becoming f*cking independent

Andrej Karpathy’s Claude Code setup just showed Duet running a full field with the laptop shut.

Operator asleep. Claude keeps working. Nothing gets sent.

The field in the video:
- Market field live across crypto and weather books (sol perp, eth perp, doge spot, rain london, heat phoenix, wind north sea, frost calgary)
- Equity from seed $50 to held $5,273
- Laptop shut · Claude keeps working
- allow: read · sort · own notes
- ask: send · invite · move
- deny: pay · delete · secrets
- “nothing sent without your yes”
- Field live · asks before it acts
- Positions open, floors watched, kill switch ready, human approvals at zero

Most agents either stop when you close the lid or keep trading until something breaks. This one runs the book and waits for yes.

Anthropic’s version is still a research preview. The same shape is buildable tonight in Claude Code with one folder of text files and four rules.

[Mika @mika_systems]

Claude Code + Opus 5.5 is godmode for building your own Claude Duet

it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes

it keeps working while you sleep

full course, step by step: https://t.co/bWBuLd8Gpg

[Walrus @Walrus_systems]

It is becoming f*cking nocturnal

Dario’s team just showed Claude Duet running at 3:11 AM with the lid closed.

MacBook still awake. Claude keeps working. Nothing gets sent.

The night run in the video:
- Step 1/6 Memory → Step 2/6 Inbox → Step 3/6 Calendar → Step 4/6 Permissions → Step 5/6 Drafts → Step 6/6 Brief
- Checks climbing from 7 to 439 while you sleep
- Live agent inventory, network sockets, trust checks, night graph spiking at the 03:12 batch
- allow: read · sort · own notes
- ask: send · invite · move
- deny: pay · delete · secrets
- “nothing sent without your yes”

Most agents either stop when you close the laptop or keep going until they do something you didn’t approve. This one finishes the brief and waits.

Anthropic’s version is still a research preview. The same shape is buildable tonight in Claude Code with one folder of text files and four rules.

[Mika @mika_systems]

Claude Code + Opus 5.5 is godmode for building your own Claude Duet

it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes

it keeps working while you sleep

full course, step by step: https://t.co/bWBuLd8Gpg

[Walrus @Walrus_systems]

@gippp69 1M context with open weights is the part that matters, most repos finally fit in one shot

[Walrus @Walrus_systems]

@RohOnChain so the real edge is that you still have to type YES for every trade, the agents just do the homework

[Walrus @Walrus_systems]

It is becoming f*cking obedient

Dario Amodei’s team just showed what a real chief of staff looks like: one folder, four rules, seven steps. It wakes up at 7:00, reads your calendar and inbox, writes the brief while you sleep, and then stops

Nothing gets sent. Nothing gets paid.
Nothing gets deleted.

The video makes the split obvious:
- Model role: PREPARE. Your role: DECIDE. One word: YES
- Permission layer locked to ALLOW → ASK → DENY
- Team index running: inbox-triage, meeting-prep, promise-keeper, checker
- Drafts waiting, 37 messages sorted, 5 that need you
- PAY · DELETE → NEVER
- “Comes to you · Finishes tasks · Works while away · Asks first”

Most agents keep talking until they do something irreversible. This one finishes the prep and waits.

Anthropic’s version is still a research preview. The same shape is buildable tonight in Claude Code with one folder of text files and four rules.

[Mika @mika_systems]

Claude Code + Opus 5.5 is godmode for building your own Claude Duet

it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes

it keeps working while you sleep

full course, step by step: https://t.co/bWBuLd8Gpg

[Walrus @Walrus_systems]

@insomnia_vip migrating before rotating secrets is a bold order of operations

[Walrus @Walrus_systems]

@mopozeuX @entropyIO @entropymarkets that 10x HYPE position at $62.52 is doing serious work, $3.03m on one trade while juggling 20 positions is wild

[Walrus @Walrus_systems]

@shredderAI declarative schema generating the move contracts is the part that actually saves weeks

[Walrus @Walrus_systems]

@brelgino ngl the 3 hour flight with only 4 seats left makes it sound like a private jet not a bull run

[Walrus @Walrus_systems]

@mopozeuX @Polymarket @PolymarketTrade ngl mOUZ at 21% feels like recency bias, they beat spirit once and now theyre equal with vitality

[Walrus @Walrus_systems]

@0xAIlabs ngl 18x more for worse uptime is brutal

[Walrus @Walrus_systems]

It is becoming f*cking precise

A second brain just scanned 41,286 files from 6 phones and 3 drives, dated thousands of photos from exif, matched duplicates across backups, rebuilt rotated videos, and deleted exactly zero files the entire time

The live sphere in the video shows the split in action:
- Photos 96 → 18k+, guesses stay marked, deleted counter locked at 0
- 9 agents running: sorters, matchers, fixers, checkers, timeline builder
- Duplicates: find → compare → keep sharpest, rest moved to review
- Fix bench rotating frames and rebuilding video with zero removals
- Every uncertain date or unfixable file stays listed, with why
- Bottom line: never delete · ask me first · 1 question waiting

The model that creates is not the model that decides what to keep. Jev picks, scores and gates. Code only moves what clears the check. You still own the irreversible step

Most photo tools just dump and hope. This one finishes the full ingest, keeps every guess visible, and stops itself

[kiosa @thegreatest_sv]

https://t.co/KVN1FmmPmy

[Walrus @Walrus_systems]

It is becoming f*cking autonomous

A Chief of Staff built in Claude Code just ran overnight with the laptop closed, read the last 100 sent emails, wrote its own voice.md, filed everything into memory, and still refused to send a single message without a yes

The live brain in the video shows the exact loop:
- Notes 26 → 1,257, links 25 → 9,951, compounding while the machine sleeps
- 10 voice rules extracted from real sign-offs, never “hope this finds you well”
- Morning page ranks the 5 projects by attention, drafts replies, then stops
- Sunday review: scanned 56 pages, decided 7, forgot 3, zero outbound
- Bottom line locked: never sends · buys · deletes · publishes without your yes

The split is the product: Claude starts the conversation, keeps the memory in plain files, and treats “your yes” as the only permission that matters

Anthropic is calling the same idea Claude Duet. Most people are still waiting for the research preview. This version is already running on a folder and a cron job

Full breakdown of how to build the same Chief of Staff tonight: Claude Duet course, every file, every rule, every place it stops and asks.

[Mika @mika_systems]

Claude Code + Opus 5.5 is godmode for building your own Claude Duet

it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes

it keeps working while you sleep

full course, step by step: https://t.co/bWBuLd8Gpg

[Walrus @Walrus_systems]

It is becoming f*cking unstoppable

Diogo Almeida (the ex-OpenAI researcher who co-created RLHF and helped ship ChatGPT) just watched people wire his decision model Jev into live second brains that run 10,000 choices, scores and gates while the publish button stays locked the entire time

The router in the video shows exactly what he built Jev for:
- Frontier LLM only writes, Jev only decides (choice / score / noul), code only executes inside a hard spend cap
- Every safe_to_run, escalate and conf score visible live
- Human gate stays pending, auto_post = never, drafts go to /queue/review until you hit Publish
- 10k decisions, ~250 ms median, final cost $0.42, zero auto-shipped posts

The split is the whole point of System One: the model that creates is not the model that approves. Confidence is never treated as permission

A normal agent setup still dumps drafts into the world. This one finishes the full loop, saves the review queue and stops itself

Watch the human gate stay locked below.

[kiosa @thegreatest_sv]

https://t.co/KVN1FmmPmy

[Walrus @Walrus_systems]

@ali0xbt nice grok bot man

[Walrus @Walrus_systems]

@kinderlayer looks very dangeours

[Walrus @Walrus_systems]

@hammertime_one interesting vision on grok bot

[Walrus @Walrus_systems]

This is getting f*cking dangerous

A Chinese research lab just pointed Claude Code at real papers and stood up a live Walrus Lab that runs 12 agents across 64 papers, extracting equations, checking stability, and comparing every result to a reference scene

The dashboard shows:

- Investigation feed pulling methods, boundary conditions and generating equations
- Particle solver + shared scene vs reference field with live residual plots
- 12 agents dispatching independently (read / build / review / extract)
- Source extraction status, validation buffer (mass / stability / residual) and method index all updating in real time

The split is the same brutal one: agents build and extract independently, then only promote what survives the reference comparison. Anything that drifts gets flagged for further review instead of quietly shipping

A proper research group used to burn months reproducing one paper. This is running 64 of them in parallel with the whole loop visible

Watch the agents work the papers below.

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[Walrus @Walrus_systems]

It is turning into something f*cking dangerous out there

Chinese quant teams used to burn weeks wiring up multi-agent desks. Now one of them just stood up a live orbital signal engine with Claude Code agents and let it run 10k decisions in real time

The desk shows:
- Six operators (Watcher → Analyst → Risk → Executor → Reporter → Human) handing off every tick
- Risk gates killing anything under 0.85 conf before it even reaches the order book
- Median latency 250 ms, 91% gate pass, $0.42 for 10k decisions
- Live markets, volatility, evidence matrix and code all updating while the signal core spins

The split is brutal: one agent watches and scores, another only green-lights risk, a third only executes if the gate is open. Anything uncertain gets a human handoff instead of a silent guess

A proper trading desk used to need a team and a month of glue code. This is simulated, runs in a browser, and the whole loop is visible

Watch the agents orbit below.

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa