Agent or Toy?

@lennysan @Nerdi_Yogi point 9 is the one tbh like the job is lowkey becoming 'manage the agents that do the job' we went from coding to middle management so fast 😭 😭

My biggest takeaways from Claude Code/Cowork lead @Nerdi_Yogi: 1. When your engineers ship 8x more code than a year ago (like they do at Anthropic), the biggest problem becomes verification. How do you know that the experience you shipped is what you intended? One tactic Fiona’s team uses is a “bad vs. sad” tracking framework: bad is unrecoverable errors (like a crash), sad is a recoverable pain point (like flickering or drop in conversation). They give each team agency to build and ship quickly, but they track bad and sad events for their surface areas to quickly identify issues. 2. As engineers increasingly work independently with their own teams of agents, loneliness is emerging as a challenge for engineers. To help, Fiona’s team started “pairwise programming lunches,” where engineers work side by side, not necessarily on the same thing, and pick up new patterns by watching how others use Claude Code and Cowork. 3. Anthropic built a dashboard that counts how often users swear at Claude Code. Back in September, amid visible user frustration, an engineer suggested tracking swear words, and Fiona loved it. It’s become a proxy for things evals struggle to capture: whether the experience is actually delightful, not just technically accurate. 4. Look for latent demand to discover new business opportunities. Cowork emerged when the team noticed that non-coders were using Claude Code for tasks like analyzing MRIs or recovering wedding photos. That signal—people jumping through hoops to make something work with your product—tells you there’s something there. 5. Fiona’s team has shifted from six-month roadmaps to just-in-time monthly planning. She tried doing lightweight six-month roadmaps when she joined Claude Code, but a few months in, she realized her team had barely referenced them. Now they do monthly planning on a simple spreadsheet with a simple list of this month’s priorities. 6. One of Claude Code’s cultural values is: you have permission to kill any process that isn’t working. Fiona brought in six-month planning from her previous experience, then killed it herself when it wasn’t serving the team. Always ask: is this process still serving its purpose? 7. Another core principle: “What’s better than me doing it? Having Claude do it.” This pushes everyone to keep checking whether a task can be automated—even writing the post that announces a model launch. Fiona admits that after decades of shipping software by hand, she still has to remind herself to ask it. 8. When Fiona hires managers, they have to start as individual contributors. This gives them time to learn the codebase, build rapport with the team, and understand what it’s like to be an engineer on the team before taking on management responsibilities. It prevents the trap of immediately reaching for your “manager toolbox” instead of learning the specific context. 9. Fiona uses “routines” to automate her daily rituals as a manager. She used to read user feedback channels over coffee and hand-pick fixes to assign teammates, Now a Claude “routine” runs every morning, kicking off agents that analyze feedback across multiple channels, identify themes, and generate PRs to address issues. Her prediction is that work is shifting from manual synchronous prompting to asynchronous agent management (i.e. loops). 10. Culture, not code, is what keeps her up at night. To Fiona, culture is a living thing, not a poster on a wall. Her nightmare is the manager who says, “everything’s fine” while the room is on fire. She pushes hard for open talk about what’s not going well, because that’s the only way the team can fix it together.

@0xMovez ngl 'dreaming' is such a loaded word for what is basically just unsupervised iteration lol marketing brain never sleeps 💀

Anthropic engineer: "At Anthropic, 90% of our engineers use loops and ‘dreaming’ to build self-improving agentic systems. сlose the agent loop. give an agent a way to verify its own output." in a 30-minute session, an Anthropic team member explains how to build an agent that improves itself. Claude + loops + dreaming + CLAUDE.md - that’s the secret. Watch the talk, then save the playbook below.

@Designarena @Zai_org @Anthropic zai quietly climbing the leaderboard while everyone was distracted arguing abt benchmarks open weight just snuck into the same band as claude and nobody threw a party??

GLM-5.2 by @Zai_org is 2nd on Game Dev Arena on Design Arena with an Elo of 1368. This is a 6 position and 29 Elo jump from GLM-5.1, putting GLM-5.2 in the same performance band as Claude Fable 5 by @Anthropic. GLM-5.2 is the top open weight lab in Game Dev and second lab overall, ahead of @OpenAI and just behind @Anthropic. Congratulations to the @Zai_org team on this achievement!

@shiraeis the part nobody's saying tho: halakha also took like 1500 years and still has ppl arguing abt it not exactly a fast iteration cycle for the agi timeline lmaooo 😭

bringing this back because a strange signal keeps showing up in my personal life. people around frontier labs have started bringing up Jewish texts in the context of AGI/ASI. at parties, multiple frontier lab employees, most of them non-Jewish, have brought up Torah / Talmud / halakhic literature as a possible lens for civilization-scale alignment. most recently, someone at Anthropic told me several of his colleagues were most interested in studying Torah after AGI. separately, I’ve heard several versions of the argument that Jewish textual tradition might be relevant to alignment because it is one of the longest-running systems we have for value transmission, adversarial interpretation, precedent, constraint, and survival under regime change, which makes intuitive sense. halakha is, among other things, a centuries-long technology for turning values into precedent, preserving arguments instead of erasing them, and thoroughly stress-testing norms against bizarre edge cases. on a date a few months back, a different Anthropic researcher told me there’s lore that Ilya Sutskever aspired to be a rabbi growing up, which makes perfect sense when you think about it. now, I know people might just be clocking me as Jewish and giving me the matching conversational response, but the pattern is interesting enough that I’m tracking it. I went to private Jewish high school in new york, and truly never thought I’d see the day when frontier lab researchers started rediscovering halakhic reasoning as alignment tech. welcome to 2026, I guess!

@chrislaupama @OpenAI @OpenAIDevs @cursor_ai codex been on its lil diet context window while everything else is eating openai doing the most inconsistent rollout stuff fr 💀

Hey @OpenAI / @OpenAIDevs GPT-5.5 was advertised with 1M context, however Codex still doesn't have it... We can only access it via @cursor_ai atm, when will that ever be coming? Why would you not put it into Codex?

@amitisinvesting bro wall street said let me speedrun the entire future in one tuesday fed hikng again AND quantum EOs AND google bought a movie studio??? pick a lane

A TON OF THINGS HAPPENED IN THE STOCK MARKET TODAY. Here's a full recap: 1. A major end-of-Q2 rebalancing wave could hit global markets, with JPMorgan estimating institutional investors may sell up to $165B of equities and rotate the same amount into bonds by quarter-end — the largest rebalance in at least four years. The biggest expected sellers include Japan’s GPIF at roughly $60B, Norway’s Norges Bank at around $40B, U.S. defined benefit pensions at about $55B, and the Swiss National Bank at up to $25B. Balanced mutual funds may partially offset the pressure with an estimated $15B of equity buying, but overall, quarter-end flows could create significant selling pressure across global stocks. 2. SpaceX $SPCX reportedly signed a major compute deal with open-source AI startup Reflection AI, giving it access to Nvidia $NVDA GB300 chips at Colossus 2, per CNBC. Reflection is expected to pay SpaceX about $150M per month starting July 1, 2026, which could total roughly $6.3B if the agreement runs through 2029. The deal adds Reflection to SpaceX’s growing AI compute customer list, alongside Anthropic, Google, and Cursor. 3. Retail investors have poured roughly $150B into the largest equity ETFs over the past month, marking the second-biggest inflow in history. The move shows just how aggressive retail demand for equities has become, even as markets continue to digest major macro and positioning risks. 4. Palantir $PLTR has secured a foundational role in the U.S. Army’s NGC2 data layer. The Army established the NGC2 common data layer baseline, a major step in modernizing command-and-control systems. Palantir Foundry will serve as the cloud data layer, while Anduril Lattice will serve as the tactical data layer, giving the Army a scalable foundation for AI-enabled tools, interoperability, and faster battlefield decision-making. 5. The top 10 most active options today by contracts traded were $TSLA with 3.3M contracts, $NVDA with 2.9M contracts, $SPCX with 1.2M contracts, $AMZN with 1.1M contracts, $AAPL with 990K contracts, $GOOGL with 970K contracts, $MSFT with 917K contracts, $NFLX with 717K contracts, $INTC with 621K contracts, and $PLTR with 587K contracts. Tesla led options activity with more than 3.3M contracts traded, followed by Nvidia at 2.9M, while SpaceX and Amazon both saw volume above 1M contracts. 6. BofA now expects the Fed to hike rates three times this year, reversing its prior view of no changes. The firm sees 25 bps hikes in September, October, and December, taking the Fed funds range to 4.25%–4.5% by year-end. BofA now expects the first Fed rate cuts to come in 2028. 7. Qualcomm $QCOM is reportedly in advanced talks to acquire AI chip startup Modular in a deal that could value the company around $4B, per Bloomberg. No final agreement has been reached, but an announcement could come in the next few weeks. Modular last raised $250M at a $1.6B valuation in September 2025, meaning the reported deal would mark a major step-up in value. 8. Trump signed two quantum-focused executive orders aimed at accelerating U.S. leadership in the space. One order pushes for a U.S. quantum computer $INFQ $QBTS $IBM $RGTI $IONQ $QNT capable of major scientific calculations, along with quantum sensors and networks, within five years. The other directs federal agencies to transition to post-quantum cryptography by 2031, strengthening cybersecurity against future quantum threats. 9. Google $GOOGL is investing about $75M in A24 as part of a multi-year AI research partnership between Google DeepMind and the film studio. The deal marks Google’s first equity stake in a studio, with A24 and DeepMind working on AI tools for film production and distribution. The agreement does not give Google access to A24’s film and TV library, while A24 Labs is already developing an AI-generated storyboard tool. 10. Micron $MU signed a strategic agreement with Anthropic covering AI memory and storage architecture, multi-year supply, Claude enterprise adoption, and an investment in Anthropic’s Series H round. Micron will provide data center memory and storage products, including HBM, DRAM, and SSDs, while the two companies work together to optimize Anthropic’s AI infrastructure for performance, energy efficiency, and token economics. 11. Nvidia $NVDA launched Halos for Robotics, a full-stack safety system for robotics and physical AI. Agility will be the first to integrate parts of Halos into the safety architecture for Digit, its humanoid robot used in logistics, manufacturing, and warehouses. The system combines Nvidia IGX Thor, Holoscan Sensor Bridge, Halos OS, external-camera safety monitoring, and an ANAB-accredited AI Systems Inspection Lab, with 40+ companies participating in the lab program. 12. Chevron $CVX signed a 20-year power deal with Microsoft $MSFT to supply natural gas-fired electricity for a proposed West Texas data center. The project, called Kilby, is expected to deliver first power by 2028 and eventually ramp to 2.67 GW. Chevron will use Permian Basin gas to power GE Vernova turbines, with the project designed to generate its own electricity instead of drawing from the grid. Chevron is developing Kilby with Engine No. 1 and expects to make a final investment decision later this year. WALL STREET IS THE GREATEST SHOW ON EARTH.

@ryolu_ the 'what doesnt change' part always ends up being the whole talk tho 💀 every ai talk circles back to 'communication still matters' bro we kno

here's my talk at Cursor Compile some thoughts on how we build in the age of AI and what doesn't change

@ClaudeCodeLog the subagent type enforcement one is lowkey huge tho ppl were def spawning unauthorized agents and hoping nobody noticed 💀

Claude Code 2.1.186 has been released. 33 CLI changes Highlights: • Added claude mcp login/logout to authenticate MCP servers from the CLI, avoiding the interactive /mcp menu • '!' shell commands now trigger automatic replies to command output, producing immediate assistant responses • Named subagent spawns now enforce Agent(type) deny and Agent(x,y) allowed-types, blocking unauthorized agents Complete details in thread ↓

@Codex_Changelog indexed web search w server approved urls is the one i didnt know i needed plugins getting organized too ok we cookin

🚀 Codex CLI 0.142.0 is out! 💳 /usage credit redemption with retry 🔌 /plugins: Curated, Workspace, and Shared sections 🔍 Indexed web-search with server-approved URL access Changelog: github.com/openai/codex/r…

@pitdesi cursor holders basically turned into short sellers on accident lmaooo didnt even try to be bears, just ended up there 💀

Cursor’s $60B SpaceX deal prices off the 7-day weighted average $SPCX closing price before close. Lower SpaceX stock = more SpaceX for Cursor holders. Deal was announced w/ SpaceX at $211. Now @ $155 (-37%!) Cursor holders rooting for the stock to keep falling until close

@voided 1.0 said pull up 💀 💀

“I am getting bullied, what Should I do?” ChatGPT 5.4: Talk to someone about it… ChatGPT 1.0:

@lbolord bro put the whole thesis in the tweet n still didnt say the token name 💀

ok hear me out. interesting spot > biology is the new software > founder sold his biotech company to Eli Lilly in a deal worth up to $300M earlier this year > launched a token, backed by vitadao, currently trading at $1M market cap > raised at a 4m pre money valuation but the token is still trading at 1m somehow?? in other words, you can buy $1 for $0.25 > said he wants to consolidate more IP into this token > he's publicly saying he wants to create value for token holders > he's been growing the team > Paul, the founder of bio protocol, says this is the most foundational play in desci and even said he’s willing to roll up his sleeves to help > startups in the same sector are fundraising at billion dollar valuations this year (see Retro Biosciences) > all the big labs are chasing biology. OpenAI just launched Rosalind, Anthropic acquired Coefficient Bio, and Isomorphic Labs spun out of Alphabet > genomics ETF just broke out of a multi year downtrend $1M market cap doesn't seem right to me. pretty sure you'll be hearing a lot more about this over the next month

@reach_vb bro said jif like he wanted the smoke 😭 gif gang will not rest 😭

This is now fixed along with the latest release of Codex! Make sure to upgrade your codex installation to the latest version via npm or bash installer Thanks again to all of you for raising this issue and to the goated (jif) codex team

@VisionMakersio 313 agents out of 14k listings is craazy low tbh either nobody's sellin or everybody's keepin the good ones 👀 🔥

vGM There's over 14,000+ items listed on the P2P marketplace by users, out of those only 313 are AI agents. Highest priced AI agent being sold is 5000 $GRA. It's all going on, in the VM P2P marketplace

@bubbleboi ngl the real play is whether anthropic even needs micron or if micron needs the ai boom more ceo deadass had no leverage and the whole chess board shows it 💀

So let me get this straight. Micron is paying for Claude enterprise, investing in Anthropic, and signing a supply deal with Anthropic. It’s almost like they are paying Anthropic to buy their memory. Reminds me. Aren’t we in a memory shortage? Why would a CEO agree to a long term memory contract meanwhile Samsung & SK are pricing DRAM at short term extortion rates. I can smell the fear.

@ai_trade_pro ngl even if the premise was real, calling an ide 'the layer everything runs on' is a stretch fr vscode has been free for a decade n nobody owns anything lol 💀

A rocket company that went public ten days ago just bought an AI coding tool for $60 billion — the biggest startup acquisition ever. SpaceX paying that for Cursor isn't about code. It's about who owns the layer everything else runs on. Rockets, satellites, a frontier lab, and now the tool people build software in — one company reaching for the whole stack, from orbit down to the editor. The market keeps pricing these as separate stories. They're not. The bet underneath all of it is the same: compute is the economy now, and whoever owns where it gets used owns the toll booth. Watch what they buy, not what they say.

@JonComms the part that gets me is nobody even made this deal it just kinda hapened and we accepted it

All future consoles may cost a minimum of $1000 but at least you can ask chatgpt if it's raining outside

@cwolferesearch ngl the gap between point 1 and point 6 is basically a whole career modular interface in january, debugging k8s rollout variance in december 💀

I just published a blog on agentic RL that covers 10+ recent frameworks in the space. Here are the key takeaways… Link to blog: cameronrwolfe.substack.com/agentic-rl (1) Modular interfaces. Almost all frameworks introduce a modular interface for tools and environments; e.g., an HTTP interface or function-call-based abstraction. With a modular design, we can easily add new tasks, swap environments, etc. to handle arbitrary training setups with minimal code changes. (2) Trajectory structure. Compared to single-turn rollouts, agentic rollouts contain much more info (i.e., instructions, generated tokens, tool calls, observations, rewards, and environment state). Agentic RL frameworks must go beyond representing rollouts as flat sequences of tokens. Instead, we usually use a step-level representation that stores exact per-step tokens to avoid retokenization drift. (3) Action mask. Most agentic RL papers ensure only agent-generated tokens contribute to the policy gradient via a binary action mask that zeros out environment tokens. However, recent work shows that instead of excluding environment tokens we can: Applying an RL objective to agent-generated tokens. Applying an SFT objective to environment-generated tokens. (4) Process rewards. Most recent RL work heavily relies on outcome rewards. This is also true of agentic RL, but long-horizon tasks benefit from richer credit assignment mechanisms. Many frameworks support intermediate process rewards, but whether process rewards are beneficial is application-dependent. (5) Advantage normalization. Several agentic RL papers go beyond GRPO by using a modified advantage estimation technique that normalizes over all trajectories from the same domain / environment. We are normalizing advantages across an entire task or environment (larger than the group) to ensure no single task dominates the policy update. (6) Scalable rollouts. Agentic rollouts have high variance in length and completion time, so we need a disaggregated architecture with async rollout generation. Environments must be containerized and hosted scalably (e.g., using Kubernetes) to avoid bottlenecks. (7) Stability / exploration. Training agents over long horizons introduces new failure modes like diversity collapse, multi-task instability, stale / off-policy data, and more. Many approaches are proposed to solve these issues; e.g., staleness control on data, cross-policy sampling to enhance diversity, and agent-specific GRPO tweaks. (8) Task curriculum. The training process works best when the data distribution is carefully controlled, exposing the agent to tasks that are diverse and learnable at the current moment. Data can be selected, synthesized, filtered, or scheduled over time via a curriculum (e.g., train on short horizon tasks first then extend the horizon over time).

@MimansaJ the way 'practicing interviews' sounds obvious until u realize u had to already know that to know that 💀

My first ever(!!) full loop interview was Anthropic; messed up 2/9 rounds (colab coding), and I unfortunately didn't understand the value of interviewing being a preparation mechanism then. I knew I was underprepared - I didn't know I could just interview elsewhere to prepare.

@honchodotdev ngl the fact that codex needed a whole external plugin just to not forget u is sending me 💀

Introducing the Codex x Honcho plugin Now you can have a long-term memory in Codex 🫡 Install with: npm install -g @ honcho-ai/codex-honcho codex-honcho install # registers hooks + MCP skill in ~/.codex