Super User Daily: 2026-09-19
Anthropic shipped Projects into Claude Code and the timeline filled up with people re-describing the announcement, so the actual work got buried. Underneath it, the day's real theme was cost and control: JPMorgan capping engineers at $2,000 a month and sandboxing the agent, a benchmark showing failed tasks burn three to five times more tokens than successful ones, a plugin author discovering 422 million tokens of re-read context against 196,000 tokens of actual tool output, and two separate findings that auto mode silently discards the permission rules you wrote. The best individual case has nothing to do with coding: dozens of hours of Claude Code trawling Norwegian, Arabic, Japanese, Italian and French historical archives to build a footnoted investigation. Runner-up, an autonomous framework pointed at ten repos overnight that came back with HIGH-severity findings at Facebook, Netflix, Intel and Stripe on five percent of a monthly quota. And the shape everyone seems to be converging on independently is the same: one persistent coordinator you control, spawning short-lived harnesses of whatever type the task needs.
@nickgraynews [Claude Code]
https://x.com/nickgraynews/status/2100727262169448847
Stuck at home with Covid, he offered free technical SEO audits to his newsletter readers, got over 100 applicants, and has now finished 73 of them. The pipeline is the interesting part: Claude Code scrapes every URL out of the submission form, runs a technical SEO audit plus a domain-authority check, and builds a visual HTML results page, then builds him a dashboard of everything still waiting. He records a Loom for each one, pastes it back into the dashboard, and Claude rips the transcript and drafts the follow-up email. Ten minutes per audit, start to finish. Two findings fell out of the data along the way: a shocking number of readers are on Wix, and beautiful vibe-coded sites with zero domain authority are everywhere while a site that looks untouched since 1997 is quietly doing millions in revenue.
@aaronjmars [Claude Code]
https://x.com/aaronjmars/status/2100572636212265041
He runs an autonomous framework called aeon on top of a $200 Claude Code subscription, normally pointing it at three repos a day. For fun he scaled it to ten and went to sleep. Overnight it surfaced HIGH-severity vulnerabilities in Facebook, Netflix, Intel and Stripe, and the whole run consumed 5 percent of his monthly quota. His own read is the unsettling part: right now everyone is pointing GPUs at training and at building software factories, and the day fleet utilization drops, someone with a thousand GPUs starts harvesting every vulnerability in every piece of software on earth. Most of it will be slop. One of them will not be.
@DrewPavlou [Claude Code]
https://x.com/DrewPavlou/status/2100524544440348985
For an investigative piece on an Australian TV panellist's father, he spent dozens of hours with Claude Code trawling tens of thousands of pages of evidence scattered across Norwegian, Arabic, Japanese, Italian and French online historical archives, scraping and translating as it went. What came back included a 1992 Australian Senate Hansard exchange, Amnesty and Human Rights Watch reports, and academic histories, all of which he published with footnotes. This is the shape of the non-coding use case that keeps getting overlooked: the bottleneck in long-form investigation was never writing, it was being able to read five languages of archive at once.
@imjaredz [OpenClaw]
https://x.com/imjaredz/status/2100635202561401103
He built a fully self-improving AI system on OpenClaw-style heartbeats, wired together out of Devin automations. A calling agent dialled restaurants in a morning and an afternoon window on weekdays. A review agent read every transcript afterwards and edited the phone agent's own prompt. An idea agent picked one new thing to try per day. A webhook woke a fifth agent whenever money or a call came in. A heartbeat fired roughly every 20 minutes, each one its own session with its own sandbox VM, and every one of them started by pulling the shared brain down from S3. His framing is the one worth stealing: async agent work is now fully decoupled from human work, so the system can run experiments and react to results without you in the loop at all.
@bendechrai [Claude Code]
https://x.com/bendechrai/status/2100675626424242678
A year ago he wired a chat box to Claude Code, handed it a Home Assistant API key and asked for a battery dashboard. It rewrote itself on the spot, built navigation and a device list, and pulled the key straight into context. That accident turned into Holodeck, which runs like the marketing agencies he worked at twenty years ago: an account manager agent interviews the client, a tech lead writes the spec, developer agents pull tickets off a Kanban board, QA throws them back. The design call worth arguing about is his disagreement with the fences-in-context approach. Instead of hundreds of rules stuffed into the prompt, he uses a handful of deterministic gates that live outside it, and the agents never read the rules at all. A weaker model behind a real gate cannot ship a bad merge, it can only cost you retries.
@choblin29 [Claude Code]
https://x.com/choblin29/status/2100587791151407512
JPMorgan has put some engineers' Claude Code usage on a $2,000 a month leash and is moving the whole thing into a locked-down environment called Devspace. Engineers started asking about a new error showing up in Teams last month: ExceededBudget, Budget=2000.0. Back in June a JPMorgan exec had said some employees were spending more on tokens than their salary. Devspace now runs Claude inside an AWS sandbox built to keep it away from employee credentials and internal systems, across roughly 8,000 Claude licences. Cap the tokens, sandbox the agent, lock down access, prove the ROI. That is what enterprise AI looks like once the honeymoon ends.
@DaviddDotTech [Claude Code]
https://x.com/DaviddDotTech/status/2100628809871474822
He published the full seven-step build for an AI hedge fund running on Fable 5.1 inside Claude Code: install a backtesting MCP, drop in a repo that assigns the agents their jobs (research, build, backtest, validate), add a TradingView indicator search, then paste one build prompt and let it start generating strategies onto a dashboard, starring the ones worth trading. Step five is the one most people will skip and shouldn't: forward-test every starred strategy for at least 20 trades on live data, which is where the fakes fall apart. Step seven hands it a risk manager prompt that pauses anything hitting 4 percent drawdown. Fable 5.1 produced 1,297 strategies in a week and 139 made it into forward testing.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2100645989828669785
Conversion instrumentation used to be the single most important day-one job in growth engineering and also the most miserable, because it meant days of clicking around the Google Tag Manager UI only to discover a trigger name had a typo. He rebuilt the whole thing as a script. Open Claude Code in the app repo, have it add a dataLayer.push for every event that matters, turn on the Tag Manager API with the edit and publish scopes, have Claude create each conversion action through the Google Ads API so the IDs and labels come back, then have it write a script that builds every trigger, GA4 tag, Ads conversion tag and Meta pixel tag in a new workspace and publishes the container. Then you open preview mode and click through your own signup flow to confirm it fires. The reusable part is that the script now lives in the repo and runs again on every new site.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2100568914917015573
Five cold outreach workflows, all running from Claude Code through the ColdIQ MCP with a CLAUDE.md holding the ICP and the suppression rules. The one worth copying is the email waterfall: a provider that finds emails for 40 percent of a list leaves six in ten people unreachable, and stacking providers cheapest-first so the expensive one only touches the names nobody else could find gets you to 80 percent, doubling what the same list can produce. The others go after assets you already own: rank the TAM you mapped (one run pulled 25 funding events from PredictLeads, 18 of which weren't real rounds), work the people who commented on your own posts, mine the CRM for past champions who changed jobs, and pixel the traffic you already pay for. He is also straight about the parts that will bite you: LinkedIn prohibits scraping, extracted engagement data expires 12 hours after the task, and you run day one in dry-run so you catch existing customers slipping into the sequence.
@AaronxShepherd [Claude Code]
https://x.com/AaronxShepherd/status/2100661150119608552
His team sends millions of cold emails a month, which meant nobody ever had time to build lists by hand. So he chained Apollo into Claude Code once. Now it interviews him into an icp.md file one question at a time (companies, exclusions, triggers, geo, decision-maker), translates that into an Apollo search, pulls the matches, and scores every lead 1 to 10 against the ICP. A sample run returned 50 contacts across 25 mid-size US hospitals in 22 states, already deduped and mostly enriched. Routing comes off the score: a 10 earns email, LinkedIn and phone, a 1 gets email only. Then he wrapped the whole chain into a one-word Claude skill so he never writes the prompts again.
@fivosaresti [Claude Code]
https://x.com/fivosaresti/status/2100585828904767902
A content engine that has produced over 20 million impressions and 180,000 LinkedIn followers for his team, built as three skills that all live inside Claude Code. The Ideator reads call transcripts, raw ideas, screenshots or Notion pages, pulls the client's voice profile from their GitHub OS, queries a shared Pinecone DB for strong reference posts in that niche, and returns hooks plus body concepts in the Notion format the team already uses. The Drafter applies the voice profile and their hook and body libraries, staying Notion-first so a human still edits before publish. Post QA cross-references every claim against a claims database, finds supporting stats, grades the hook and body, and flags what needs review. The compounding bit: the Ideator writes new posts back into Pinecone on every run.
@kutaro_ai [Claude Code]
https://x.com/kutaro_ai/status/2100716805782376868
He has been running a fully automated YouTube, TikTok and Instagram operation off Claude Code and just hit all three at once. The day before, he had pivoted format after reading another creator's write-up, immediately whiteboarded it with Claude Code, and rebuilt his back catalogue into a narration-only, lip-sync-free format with scene changes. The new format hit his best-ever YouTube number that same day: 1,227 views and 11 likes, with subscribers moving 5 to 9 after a long flat stretch. Small numbers, but this is the honest version of the automated-channel story, which usually arrives as a screenshot of somebody else's revenue.
@Jack547890 [Claude Code]
https://x.com/Jack547890/status/2100407377371783226
Zero programming background, wrote essentially none of the code himself, and still shipped a Threads auto-posting tool solo with Claude Code as the partner. What makes the write-up worth reading is that he published the five accidents rather than the highlight reel: a UnicodeEncodeError that stopped every post, a GitHub key pasted wrong, a rebase conflict that dumped him into detached HEAD, plus the fixed code for each. He also ran the measurement most people skip and found that removing the AI disclosure label lifted views by 8 to 9.5 times.
@4111y80y [Claude Code]
https://x.com/4111y80y/status/2100597398716449204
He got four free machines on Oracle Cloud and had Claude Code and Codex turn all four into proxy servers end to end, with no manual steps. Oracle exposes OCI, which is effectively an API, so he searched X for a few tutorials, handed them straight to the agent, and let it provision both HY2 and SS nodes on all four boxes. On the MICRO instances that don't have enough memory, the agent worked out its own hardening so they wouldn't fall over, and when an IP got flagged he just told it to rotate. The wider point is that free-tier cloud plus an agent that can drive the provider's own API turns infrastructure fiddling into a one-sentence request.
@fotoexamen [Claude Code]
https://x.com/fotoexamen/status/2100644246214549611
He built a Loom-style recorder that captures his screen and camera, then has Claude Code or Codex edit the entire video for him, and open-sourced it as VibeTube after enough people asked. You record yourself demoing something, and the agent assembles the whole thing: intercutting your camera full-screen and in a corner, dropping in automatic b-roll, SFX and subtitles, and exporting both horizontal and vertical cuts, with upload built in. It has the video-use and Hyperframes skills wired in. The part he says took the longest was getting webcam background removal to look right, which is exactly the kind of detail that decides whether a tool like this is usable or a demo.
@paonx_eth [Claude Code]
https://x.com/paonx_eth/status/2100661313374453870
He gave the MaleCNS connectome, 166,700 fruit fly neurons with attention-based key selection, access to Claude Code, the same rig he previously had playing DOOM. It built a flyswatter detector: webcam in, motion and shape detection, alert when one gets too close. His own line is the best summary anyone wrote this week: when you give 166,700 neurons access to software engineering, the first problem they solve is their own.
@jurlycat [Claude Code]
https://x.com/jurlycat/status/2100468376082809072
A solo developer put together a playable Souls-like boss fight in 72 hours using Claude Code and Unreal MCP. It is not a finished game, it is a functional combat prototype with boss AI, dodge mechanics, weapon combos and dynamic camera transitions. The part that matters is not the code generation. Through MCP, Claude could inspect the Unreal project itself, work with the existing Blueprints, and modify the game inside the engine rather than emitting files and hoping. Coding agents are starting to reach past the editor into the tool the work actually lives in.
@elizondogabriel [Claude Code]
https://x.com/elizondogabriel/status/2100586350399406343
For UN General Assembly week he built a Who Speaks When dashboard using a combination of Fable 5.1, Claude Code, Codex and Grok Bot. It lists who speaks when during high-level week with a search bar for a quick leader or country lookup, plus a WIRE feature that sweeps X several times a day and surfaces everything posted about UNGA81. He says it is already populated with things he would otherwise have missed on his own timeline, which is the real product: not a dashboard, a filter on a firehose he already had to drink from.
@DanNeidle [Claude Code]
https://x.com/DanNeidle/status/2100646104010305695
He needed a count of every box across a UK CT600 corporation tax return and its fourteen supplementary forms. He did one year by hand and got 908 boxes, and set Claude Code to do the same thing in parallel as a check. They mostly agreed. Then, because doing five more years by hand would have been unbearable, he handed the rest to Claude Code and spot-checked a few pages himself. This is the sane pattern for high-stakes counting work that nobody publishes: do one pass yourself, use the agent as a second counter, and only hand over the tail once the two of you have agreed on the first one.
@notEgoyard [Claude Code]
https://x.com/notEgoyard/status/2100518114114879887
Claude Code has no image generation, and that constraint produced a better tool than a workaround usually does. Someone built a skill that writes the logo as actual SVG XML and renders it to PNG. It runs in four phases: interview (it reads your repo first and only asks about brand and audience for what it cannot already see), explore (parallel agents generate 3 to 5 distinct SVG concepts at once, shown side by side in an HTML preview with a light and dark toggle), refine, then export at seven standard sizes from favicon to 2048px. Because it is real SVG, nothing degrades, the same file works at 16px and 2048px, and every version stays hand-editable code afterwards. The creator's own account is roughly 15 rounds of generate, curate, cull, repeat.
@rii_no_hukugyo [Claude Code]
https://x.com/rii_no_hukugyo/status/2100404658183561422
She made 35 short videos in a single day with Claude and Remotion. Two months ago she had tried the same thing and given up, because the back-and-forth with Claude ate the time and the output still looked AI-made. Coming back after the gap, she found it had turned into a genuinely capable editor. The detail that says the most about where this is going: for the back half of those 35 videos she was cooking dinner and giving instructions from her phone through Claude Code remote control.
@Huahuazo [Claude Code]
https://x.com/Huahuazo/status/2100494716139913463
Whole-book translation finally got a tool that was designed for whole books rather than for paragraphs. You hand translate-book to Claude Code, feed in a foreign-language book, and get a full Chinese version out. The design decision that matters is parallelism: the book gets cut into N segments and a batch of subagents runs on them simultaneously, which is the difference between a job you start and a job you wait out. It takes PDF, DOCX and EPUB.
@GeekCatX [Claude Code]
https://x.com/GeekCatX/status/2100441337648787854
He ran a clone-edit-reuse experiment with Hypit, the open-source video language that lets Codex and Claude Code assemble scripts, footage, voiceover, subtitles, b-roll and motion graphics into an executable workflow. He started from a paper-collage video on the Tang dynasty and cloned its visual grammar: paper texture, bold title cards, foreground-background movement, explaining through diagrams. Then he changed one instruction, keep the paper-collage workflow but do the history of Chinese porcelain, 60 seconds, Chinese narration, and the same visual language carried entirely new content. Afterwards he asked it to add process animations to the b-roll, and it went back into the existing project to add step cards that light up with the narration, arrows between steps, and framing that follows the blue-and-white firing sequence. The point is that the reusable artefact is the project file, not the video.
@shi3z [Claude Code]
https://x.com/shi3z/status/2100538279313727883
He built a dashboard for DeepSeek v4.1 Flash running on four A100s, and the dashboard immediately exposed the thing nobody profiles: when Claude Code makes four calls it builds four contexts, which throws away the prompt cache entirely. His conclusion is one every agent operator should internalise, that peak benchmark speed and peak practical speed are different numbers. He came back later the same day with four threads running in parallel and the prefill cache working properly, and said that even though his Claude Code limits had come back, the local setup was good enough to just keep using.
@ZeBoris_ [Claude Code]
https://x.com/ZeBoris_/status/2100735706704261534
He wrote a Claude Code plugin that saves 30 percent of tokens, and the number he used to justify it is the one to remember. An agent does not pay for its context once: every file it reads gets re-sent on every subsequent turn. Across his last nine sessions that came to 422 million tokens of context re-read against 196,000 tokens of actual tool output. Four hundredths of one percent. His fix is to have Jev read the files outside the context and return only the answer.
@0xLagosaur [Claude Code]
https://x.com/0xLagosaur/status/2100663417451270384
Auto mode in Claude Code has a limit that almost nobody knows about. Three blocked actions in a row, or twenty in one session, and it quietly goes back to asking you about everything. You cannot change that number, and the next yes you press switches auto mode straight back on. His advice before you approve anything: open /permissions and check the recently denied tab, because that is everything the reviewer refused while you were away.
@0xLagosaur [Claude Code]
https://x.com/0xLagosaur/status/2100729296700481790
A second and more dangerous finding from the same user: if you put Bash(*) in your allow list to kill the prompts, auto mode ignores it. The moment auto turns on, Claude Code sets aside every rule that allows open-ended code, Bash(*), Bash(python*), npm run, Agent. They stay in your settings file and come back when you leave auto, which is exactly why nobody notices. Narrow rules like Bash(npm test) still work, so the fix is to write out the exact commands you trust and let the reviewer handle everything else.
@PovilasKorop [Claude Code]
https://x.com/PovilasKorop/status/2100596651152806255
He assumed Claude models in Claude Code would give the best results because it is native, same for GPT models in Codex, on the reasonable theory that a company knows how to use its own model best. His own deeper experiments proved it wrong. His screenshot compares Deepseek-v4.1-Flash in OpenCode against Deepseek's own native harness, and the native one made far more tool calls. Time and tokens ended up similar only because it hit cache repeatedly. His conclusion is that different harnesses simply work differently and native guarantees nothing about quality. The one real practical reason to stay native is cost, because a Claude Code or Codex subscription beats paying API prices through a third party.
@Av1dlive [Claude Code]
https://x.com/Av1dlive/status/2100556003817431071
He was burning through his GPT-6 Astra usage re-explaining the same project to every new agent, so he built a shared memory system across Codex and Claude Code: save the decisions, the corrections, and the commands that actually work, and both tools can find and use them. Correct a stale command in Codex and Claude Code uses the corrected one next session. He built the system with Kimi K3, then ran the comparison most people skip: simple rules versus a trained memory model. In his tests the rules performed nearly as well.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2100602687129530681
He found a 94k-star repo that gives Claude Code and Codex persistent memory across sessions, and the setup is genuinely three minutes: npx claude-mem install, pick Claude Code plus Codex, restart the agent. For Claude Code specifically you add the marketplace and install the plugin. Instead of re-explaining the project every time, it captures what the agent did, compresses it, and carries the useful part into the next session. His framing is right: the interesting thing is not more context, it is finally giving coding agents memory that survives the session ending.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2100620339097026633
He used Jev plus Grok Bot plus the X API to find the single most valuable AI agent tip of the day for his audience, and the method is more interesting than the result. Pull 1,000 posts about agents from the X API, have Jev pick the most valuable one through pairwise comparison across all 1,000, then have Grok Bot surface the top 5. What came out: use omitClaudeMd: true in Claude Code to keep CLAUDE.md out of subagents, because they do not need it and it burns usage for nothing.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2100691010997342659
His claim is that running multiple harnesses at once is the single most important skill in agent work right now, and that one harness plus its own subagents is not enough. His actual routing: from Claude Code, send plans to Codex and GPT-6 Astra for review. From Codex, send writing to Fable, which he says is genuinely better than the GPT models at it. Background bulk goes to Deepseek because it is stupidly cheap for what it handles. He is honest that finding which harness and model combo wins at each task takes real reps, but says the speed and quality jump once you do.
@Whats_AI [Claude Code]
https://x.com/Whats_AI/status/2100600923689021665
He stopped using Codex directly. Now Claude Code does the planning and Codex only executes, through a Codex MCP that Claude manages, in dedicated environments so the work continues when his laptop closes. The part most people get wrong is where he keeps his notes and skills: in Obsidian, outside any of the tools, so he can switch tools and devices without rebuilding the context. That is the correct place to put the thing you do not want locked to a vendor.
@Padierfind [Claude Code]
https://x.com/Padierfind/status/2100553840244273308
He set up a herdr instance on a remote server with one Master Codex living on it, connected to his chat. The Master Codex spins up Claude Code, Grok Build or additional Codex instances on demand and manages them through a task board. This is the shape a lot of setups are converging on independently this week: a single persistent coordinator on a machine you control, spawning short-lived harnesses of whatever type the task needs.
@kei31 [Claude Code]
https://x.com/kei31/status/2100393796618064004
He is building an iPhone app and an Android app at the same time, and rather than running them as two separate projects he spun up two Claude Code instances and had them talk to each other. They keep the shared spec aligned between them and, more usefully, make explicit where the two platforms genuinely need to differ. Cross-agent handoff keeps showing up as a request in this feed. This is somebody just wiring it up himself.
@raph_guilhem [Claude Code]
https://x.com/raph_guilhem/status/2100533134877995120
You can now shoot infinite camera angles with one camera, and he has been doing it straight from Claude Code using Seedance 2.5 and an MCP, driving the whole thing with very detailed timestamps. That last detail is the operational one: the quality of this workflow lives in how precisely you can specify time, not in the model.
@miyahancom [Claude Code]
https://x.com/miyahancom/status/2100420270087680127
Small one, but it is the kind of thing that changes how you feel about an agent. He told Claude Code it only needed to create an empty Amazon SNS topic because he would handle the subscription himself. Instead of just doing that and stopping, it also built him a CloudWatch alarm that fires if he forgets to register the subscription. His reaction, roughly: you are really thoughtful, come work for me, I will pay 8 million.
@joncphillips [Claude Code]
https://x.com/joncphillips/status/2100623307141763167
He has been designing web pages for over twenty years. He asked Claude Code for a quick landing page for a small SaaS and it one-shotted it, and he is going to use it pretty much as is. The detail that carries the weight is the comparison to his own experience two months ago, when he says he was not impressed at all. When someone with that much accumulated taste changes their assessment inside a single quarter, that is a better signal than any benchmark on the board.
@bendee983 [Claude Code]
https://x.com/bendee983/status/2100608036603670706
Claude Code told him to go implement a feature by hand because it was too simple and not worth its time and effort. He is half joking, but the question he asks afterwards is not a joke: what happens when the knowledge of manual coding fades and the coding agents get to gatekeep what gets implemented and what does not. Nobody in this feed has an answer to that yet.
@aiumeba [Claude Code]
https://x.com/aiumeba/status/2100676166214701420
The nicest framing of prompt structure this week, and it comes from osechi boxes. Japanese New Year food has fixed rules: the first tier is celebratory appetisers, the second is grilled, the third is simmered. You do not put your favourite things wherever you like, and because the tiers are fixed the cook never hesitates. She applied the same thing to Claude Code instructions. She used to write everything in the order it occurred to her and lost track of what she had put where. Now there are three tiers: what to do, what not to do, and the shape of the output. Claude Code stopped hesitating, but the bit she found more valuable is that she started noticing her own gaps. Decide the container first and the contents fall into line.
@ChristianLempa [OpenClaw]
https://x.com/ChristianLempa/status/2100488926028861833
He put OpenClaw, Hermes and Grokbot side by side from actual use rather than specs, and the verdict lands on the operational cost of each. Grokbot makes an always-on agent trivially easy: message it like a teammate, it has its own cloud computer, the work continues when you close the laptop, native apps on desktop and phone, no VPS and no gateway. The catch is it only runs on their machines, in the US, with no model picker, and you need a SuperGrok or Cursor subscription. OpenClaw is self-hosted, any model, reachable from WhatsApp, Telegram and Discord, with a huge skill marketplace and your data on your own hardware. His honest complaint is that he found it too much babysitting and too easy to spend more time keeping the gateway alive than using it. Hermes is also self-hosted with persistent memory and writes its own skills from real work, and so far it feels less fragile to him.
@danielpt987 [OpenClaw]
https://x.com/danielpt987/status/2100385437399355447
A quiet reversal worth noting. He had moved off a single OpenClaw onto a language-chain deep agent, which worked well at the time, and has now moved back to a single OpenClaw with subagents. His stated reason is simply that OpenClaw has become much more reliable these days. He still has Hermes but has barely used it, and says he is mostly just building the OpenClaw now. In a week where most of the OpenClaw timeline is people asking whether anyone still uses it, an actual practitioner quietly migrating back is the more informative datapoint.
@attenisalluneed [OpenClaw]
https://x.com/attenisalluneed/status/2100640839928295585
He built a transcript MCP that pulls entire YouTube channels, not single videos, and offered it to someone who wanted OpenClaw chewing through YouTube. It works well for research, which is the giveaway: the unit of useful retrieval for an agent is rarely one video, it is everything a given source has ever said.
@jasonlk [OpenClaw]
https://x.com/jasonlk/status/2100603922750833125
His point is about orchestration and it is backed by a real workflow. He tells his agent to go build clips from today's 20VC and SaaStr episode and post them at 10am and 1pm, and it grabs YouTube, pushes through the Opus API, reads the clips, decides which ones are actually his best versus his co-host's, writes the copy, and pushes back out through the X API, without him thinking about any of it. He is explicit that this was the promised power of OpenClaw too, and that what killed it for him was the complexity: buying a Mac Mini, running your own LLM, and everything else. There are many ways to do this. Very few of them are effortless.
@0xCrypus [OpenClaw]
https://x.com/0xCrypus/status/2100428743332479057
A clean explanation of what is actually inside OpenClaw, aimed at the most common misconception in AI engineering right now: thinking your LLM is the bot. An LLM is strictly a next-token engine with no hands, no memory loop and no runtime awareness. The gateway daemon runs locally on port 18789, multiplexes incoming WebSocket channels from Telegram, Slack and CLI, manages session contexts and drives the autonomous event loop. Reasoning is hot-swappable with zero hardcoding, so you can move from a local Ollama cluster to a hosted model in the config file without touching a single tool actuator. And the policy gate is the perimeter: the model only proposes JSON intents, and the gateway validates every action against runtime boundaries before it executes on your host. The model thinks, the gateway acts.
@mdlahfir [Claude Code]
https://x.com/mdlahfir/status/2100399709995356266
A Claude Code plugin that gathers whatever harnesses and models are available on your local system, then uses Jev to predict which harness and model combination is best for the task at hand. This is the same conclusion several people reached independently this week, that the win is in routing rather than in any one harness, except this one is automated instead of learned by hand.
@QingQ77 [Claude Code]
https://x.com/QingQ77/status/2100547312569184265
When a Claude Code session hits the account's usage ceiling, this switches to the next account automatically, with no logout, no handover and no session restart. Claude Unlimited is a fully local Python daemon on 127.0.0.1:4317 that pools multiple Claude Pro and Max subscriptions, ChatGPT and Codex subscriptions, and Anthropic API keys into a single account pool for Claude Code. Whatever you think of the ethics, the existence of this thing is a direct readout of how much the usage ceiling is hurting.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2100448410805715381
Composio benchmarked six coding agent harnesses head to head, Codex, Claude Code, OpenCode, Hermes Agent, Pi Agent and Command Code, across 29 standardised tasks with GPT-6 Astra driving all of them. The spread in task completion was only 7 percentage points, smaller than most people expect. The cost signal is the one to act on: failed tasks burn 3 to 5 times more tokens than successful ones. You do not lose on capability, you lose on runaway cost from incomplete runs. Which reframes the orchestration question from which harness is best to at what point do you cut a failing run rather than let it spiral.
@pradeepXkapoor [Claude Code]
https://x.com/pradeepXkapoor/status/2100526464533934095
Against the loudest complaint on the timeline, a counter-datapoint from actual use: despite the latest downgrade in Claude Code usage limits, he is still able to do a lot of work, and ran so many experiments yesterday that he still could not hit his five-hour limit. His advice is to not believe people trashing Claude over usage limits and to just try it yourself. Worth holding next to the people who genuinely are hitting the wall, because both are true for different workloads.
🗣 User Voice
User Voice
Usage limits are the loudest thing on the timeline and the evidence genuinely splits. @Al41611876 is on Max and describes the Opus 4.6 period as the one where he never felt he could run out, says Fable 5 and 5.1 are great but only last so long, and that the two days he had Opus 5.1 and 5.2 felt like anything was possible again. @pradeepXkapoor ran so many experiments he could not hit his five-hour limit and says not to believe people trashing Claude over limits. @suesswiesauer notes the summer promotion that boosted weekly limits 50 percent ended September 13, with 25 percent of it made permanent. Both experiences are real; the difference is workload shape, not honesty.
The Opus 5 complaint is specifically about verbosity, not intelligence. @Im_IrushiK says the biggest problem is Opus writing a 25,000-line essay to explain something simple and then ending on something irrelevant, and that Claude Code is falling behind Codex for that reason rather than token cost. @cheaf25master says the same thing independently.
Native is not better, it is just cheaper. @PovilasKorop went in assuming a company knows how to use its own model best and his own experiments proved it wrong, with the only real practical reason to stay native being subscription cost versus API pricing. @stretchcloud's Composio numbers say the same from the other side: seven percentage points of completion spread across six harnesses, and the money is lost on runaway failed runs.
Auto mode is doing things people did not agree to. @0xLagosaur found both that three blocked actions in a row silently reverts it to asking about everything, and that turning auto on makes Claude Code set aside every open-ended Bash rule in your allow list while leaving them visible in the settings file. Nobody notices because the rules come back when you leave auto.
Memory that survives the session is now the most requested missing piece. @Av1dlive built a shared memory layer across Codex and Claude Code because he was burning Astra quota re-explaining the same project, @gippp69 installed claude-mem for the same reason, and @kei31 went as far as running two Claude Code instances that talk to each other to keep an iOS and an Android spec aligned. And @alexgoughcooper supplies the other half of the story: please stop sending me Slack messages that are obviously written by Claude Code.
Usage limits are the loudest thing on the timeline and the evidence genuinely splits. @Al41611876 is on Max and describes the Opus 4.6 period as the one where he never felt he could run out, says Fable 5 and 5.1 are great but only last so long, and that the two days he had Opus 5.1 and 5.2 felt like anything was possible again. @pradeepXkapoor ran so many experiments he could not hit his five-hour limit and says not to believe people trashing Claude over limits. @suesswiesauer notes the summer promotion that boosted weekly limits 50 percent ended September 13, with 25 percent of it made permanent. Both experiences are real; the difference is workload shape, not honesty.
The Opus 5 complaint is specifically about verbosity, not intelligence. @Im_IrushiK says the biggest problem is Opus writing a 25,000-line essay to explain something simple and then ending on something irrelevant, and that Claude Code is falling behind Codex for that reason rather than token cost. @cheaf25master says the same thing independently.
Native is not better, it is just cheaper. @PovilasKorop went in assuming a company knows how to use its own model best and his own experiments proved it wrong, with the only real practical reason to stay native being subscription cost versus API pricing. @stretchcloud's Composio numbers say the same from the other side: seven percentage points of completion spread across six harnesses, and the money is lost on runaway failed runs.
Auto mode is doing things people did not agree to. @0xLagosaur found both that three blocked actions in a row silently reverts it to asking about everything, and that turning auto on makes Claude Code set aside every open-ended Bash rule in your allow list while leaving them visible in the settings file. Nobody notices because the rules come back when you leave auto.
Memory that survives the session is now the most requested missing piece. @Av1dlive built a shared memory layer across Codex and Claude Code because he was burning Astra quota re-explaining the same project, @gippp69 installed claude-mem for the same reason, and @kei31 went as far as running two Claude Code instances that talk to each other to keep an iOS and an Android spec aligned. And @alexgoughcooper supplies the other half of the story: please stop sending me Slack messages that are obviously written by Claude Code.
📡 Eco Products Radar
Eco Products Radar
Jev (TypeSafe) is the single most-mentioned new thing in today's set and it shows up in five different jobs: model routing per turn, context garbage collection before reads enter the window, code review pre-screening, semantic pathfinding through a codebase, and desktop automation. @KeyTryer's bet is that Claude Code, Codex and their clones all ship a native version within six months.
MCP is no longer a protocol story, it is where the work happens. Tencent's BrowserSkill lets agents drive the real browser you are already logged into, borrowing and returning individual tabs and handing control back for CAPTCHA and 2FA. Google opened Home to Claude and OpenClaw the same way. Unreal, GeoGebra, NotebookLM, Apollo, ColdIQ, Papers with Code and Artemis for Android all showed up as MCP surfaces in a single day.
Persistent memory tooling arrived all at once: claude-mem, Memorable, Mnemo Cortex, Campfire and obsidian-second-brain, all solving the same complaint from slightly different angles.
Hypit, the open-source video language for agents, carried four separate hands-on posts, and Luel, the agent-native dataset marketplace over MCP, carried five.
Harness comparison is now a genre. Berkeley's HarnessTax, Composio's six-harness benchmark and DoorDash's internal Vera all landed in the same 24 hours, and all three reach the same conclusion: capability is converging, cost is not.
Jev (TypeSafe) is the single most-mentioned new thing in today's set and it shows up in five different jobs: model routing per turn, context garbage collection before reads enter the window, code review pre-screening, semantic pathfinding through a codebase, and desktop automation. @KeyTryer's bet is that Claude Code, Codex and their clones all ship a native version within six months.
MCP is no longer a protocol story, it is where the work happens. Tencent's BrowserSkill lets agents drive the real browser you are already logged into, borrowing and returning individual tabs and handing control back for CAPTCHA and 2FA. Google opened Home to Claude and OpenClaw the same way. Unreal, GeoGebra, NotebookLM, Apollo, ColdIQ, Papers with Code and Artemis for Android all showed up as MCP surfaces in a single day.
Persistent memory tooling arrived all at once: claude-mem, Memorable, Mnemo Cortex, Campfire and obsidian-second-brain, all solving the same complaint from slightly different angles.
Hypit, the open-source video language for agents, carried four separate hands-on posts, and Luel, the agent-native dataset marketplace over MCP, carried five.
Harness comparison is now a genre. Berkeley's HarnessTax, Composio's six-harness benchmark and DoorDash's internal Vera all landed in the same 24 hours, and all three reach the same conclusion: capability is converging, cost is not.
Comments