Super User Daily: 2026-10-26
Saturday's timeline belonged to two goodbyes. Thariq Shihipar and Anthony Morris, two of the most visible people behind Claude Code and its desktop app, left Anthropic on the same day, and most of the feed was people reading tea leaves about the IPO. Underneath the noise, the users kept shipping. A solo Japanese game developer passed 500,000 yen in monthly in-app revenue across two games, with Claude Code writing and Codex reviewing every payment path. Two AI agents closed a gap in a Rule 30 prize problem from 62 down to 9, checked line by line in Lean. One developer counted 2,753 commits since mid-September and found that what they actually own is 140 memory notes. The other thread was the personal agent turning into a dispatcher: people now let Grok Bot hand the real work to Claude Code and Codex on a Mac mini, and say their quota lasts longer that way. OpenClaw's best numbers came from the enterprise side, 500 gateways at about 45 dollars per employee per month, while home users kept describing the maintenance bill of running their own agent. The best measured post of the day checked MCP Tasks against 8,043 real tool calls and found only three ran past two minutes.
@tatsu_solodev [Claude Code]
https://x.com/tatsu_solodev/status/2113845909880795367
A solo developer in Japan launched a hack-and-slash roguelite on iOS on July 31 to exactly three downloads on day one. Three months later the developer runs two games, and the last 28 days brought about 509,000 yen in in-app purchases and roughly 2,300 new players. The second game, an auto-battler called ECHO LEGION, went from first line of code to the App Store in 20 days. The split is strict: the developer does design, balance and final calls, Claude Code writes the code, and Codex reviews it, because the same AI reviewing its own work misses things, so anything touching payments or save data always gets doubted by a different model. Crash reports from players' phones now turn into a fix proposal from one AI and a review from another before the developer approves, which is how 952 code changes shipped in 30 days.
@AIOvernight [Claude Code]
https://x.com/AIOvernight/status/2114006234000998757
Rule 30 is Stephen Wolfram's famous cellular automaton, and Problem 1 of the Rule 30 Prize asks for a proof that its center column never becomes periodic. This team proved that for every q of 9 or more, the column can never settle into repeating q white cells followed by 10111, a result that was previously known only for q of 62 or more. The proof combines a new 8-cell invariant that survives any number of white steps with an existing argument in the kit, and every step is checked by the Lean proof assistant. The maintainers reviewed it and merged it as a module. The research, the proofs and the formalization were all done by AI agents, OpenAI Codex and Claude Code, which is about as clean a case as exists for agents doing real mathematics with a machine checker standing behind them.
@zihaopi [Claude Code]
https://x.com/zihaopi/status/2114050035134722350
Since September 17 Claude Code has been this developer's whole dev team, and the count is 2,753 commits. The more interesting number is what is left when the tool goes away: one AGENTS.md and 140 memory notes, all in plain files. Eighty of those notes are the developer saying either not like that or yes, like that. It is a neat way to see where the human's value now sits, not in the code but in a corrections log that any future harness can read.
@bstuartTI [Claude Code]
https://x.com/bstuartTI/status/2114120237834723407
This user has run Codex and Claude Code side by side for weeks on two real projects, and logged one day this week. Given a one-page rule sheet, Codex on Astra spun up nine helpers, told eight to wait, and the ninth never started because a rule Codex had written for itself said it was not allowed to; it later asked permission to add a rule that would stop it asking permission. Claude Code on Opus 5.5 ran two crews overnight, delivered three usable things by morning, rebuilt the one the user hated in under two hours, tracked down a slowdown by lunch and reached nine deliveries by mid-afternoon. The user's conclusion is that the difference is not intelligence: both models drown in status updates and self-reviews, but one stops for permission it does not need and the other keeps building.
@fagamericano [OpenClaw]
https://x.com/fagamericano/status/2114144839168299501
Replying to a claim that OpenClaw is too expensive, this team shared real enterprise numbers. Their OpenClaw deployment serves about 500 people with one gateway per employee, 500 gateways in total. Gateway and LLM calls together run about 25,000 dollars a month, roughly 45 dollars per employee. For that they keep full control over tools, connections and code, which is the trade most hosted personal agents cannot offer a company.
@kaibuildsagents [Claude Code]
https://x.com/kaibuildsagents/status/2113984152739651893
Claude Code's new MCP Tasks move any tool call that runs longer than two minutes into the background. This builder checked that threshold against 8,043 local agent tool calls and found only 3 ran past two minutes, 0.04 percent. But 64 percent of all waiting time went to sleep, polling output and waiting on subagents, because long jobs had already been chopped into 30 to 60 second polls. The practical takeaway for anyone writing an MCP server is to return a task ID for long jobs directly instead of making the agent poll.
@jacksunny18 [Claude Code]
https://x.com/jacksunny18/status/2114107568528445855
After Anthropic donated Agent Skills to a foundation, people argued over whether Claude Code reads the shared .agents/skills folder. This user checked three ways. The skills doc lists six load locations and every one is a .claude/skills path or a plugin folder, even with --add-dir; the changelog never mentions .agents; and issue 31005 asking for it has been open since August 2025 with 37 comments. The twist is that the spec itself names no folder at all, only what goes inside one, and the .agents/skills path is called a convention in the client guide. So the donation covers the format while the folder, which is exactly where the lock-in lives, stays each client's call.
@Yeq6X [Claude Code]
https://x.com/Yeq6X/status/2113837645143126225
As capable secretary agents keep appearing, this user says the best decision so far was making sure work can resume with the same information whether it restarts from Grok Bot, Muse, Dots or Claude Code. The trick was deliberately keeping the memory layer separate from the agent layer. That makes things like handing yesterday's Muse task to Grok Bot today actually work. The rule is easy to remember: never let your records live only inside one secretary app.
@DeepakNesss [Claude Code]
https://x.com/DeepakNesss/status/2113912028675162171
This user downloaded T3 Code and, for the first time ever, hit the five-hour limit on a Claude Max plan. It was a single session fixing a few payment bugs, prompted exactly the way the user prompts inside the Claude Code CLI, and the whole five-hour window was gone in about 30 minutes. The post is a plain request for help, but it is a useful warning: a third-party front end on top of the same subscription can burn quota on a completely different curve than the official harness.
@cagrimbakirci [Claude Code]
https://x.com/cagrimbakirci/status/2114077715058114806
This user reports that Claude Code's token burn on Opus 5.5 has been really bad over the past few days, at least two to three times the usual rate. Boris Cherny has been looking into it for them, but so far the problem persists. It is a short post, but coming from a heavy user who already escalated to the team, it is a cleaner signal than the usual anonymous complaints about limits.
@clairevo [Claude Code]
https://x.com/clairevo/status/2114012293105430813
This user says 80 percent of personal coding has moved over to Claude Code plus Opus 5.5. The reasons are concrete: Claude Code projects work well, the steer button in Codex is gone, the code is good, and there is less front-end slop. The honest admission is that Opus 5.5's creative ability was the gateway drug that pulled the coding work back. It is a nice data point that model taste, not only benchmark scores, is what moves people between harnesses.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2113842948945412240
Anyone trying AI short dramas finds dozens of drama skills on GitHub and no way to tell which ones actually write good scripts. This curated list collects 83 short-drama skills and 16 related open-source platforms, each checked by hand, organized by stage from novel adaptation and episode outlines to hooks, script review, storyboards and video prompts for Seedance, Kling and MiniMax. What makes it worth trusting is the test: the author ran 10 script skills against a no-skill control group, three runs each on the same brief, judged blind by AI, and only 5 beat the control all three times. The limits are listed honestly too, including that the episode-ending hooks were often better in the control group. It installs into any skill-capable agent, Claude Code and Codex included.
@likebookmark [Claude Code]
https://x.com/likebookmark/status/2114063238015103488
Feeding a novel straight to an AI for a short-drama script usually produces characters who change personality halfway and props that do not match, and everything has to be redone at the storyboard stage. This user ran the open-source shuohao-skills pack inside Claude Code for a full novel-to-drama pass: adaptation outline, character sheets, scene and prop specs, episode scripts, blocking and storyboards. Each stage ships with local validation scripts that block mismatched character names or wrong aspect ratios instead of trusting the prompt, and each stage outputs a single-file HTML report that can merge outline, characters, art, script and storyboard into one page for review. The caveats are clear: it does not call image or video APIs for you, and its validation rules assume Chinese output.
@mreflow [Claude Code]
https://x.com/mreflow/status/2113785431917760952
This creator edited a video for the first time with no DaVinci Resolve or similar editor, only Claude Code. The creator recorded the a-roll of talking to camera, and Claude Code sourced the b-roll, timed it to the speech, split the frame top and bottom and added subtitles. It is a small first attempt, but it is the clearest version of a workflow many people keep describing: the editor becomes a folder and a conversation.
@biscuitweb [Claude Code]
https://x.com/biscuitweb/status/2114096899477106954
This user now edits their own videos with Claude, and the setup is short. Install Claude Code, have it set up an open-source editing kit from its install guide, add an ElevenLabs key, drop the raw clips into one folder and open Claude Code there. The prompt is plain: remove bad takes and long pauses, keep the natural rhythm, add captions in my style and export 9:16. Claude transcribes the clips, shows a cut plan for approval, writes the result to edit/final.mp4, and later changes are given by timestamp, like undo the cut at 00:18.
@thesourabhbhati [Claude Code]
https://x.com/thesourabhbhati/status/2114063661249806697
This builder made a launch video, music included, in under two hours with Claude Code. The workflow is four steps: find a reference video, break down its style, prepare a storyboard, and ask Claude to render it. The tip that matters is precision: watch a lot of launch videos first so you have a clear picture of what you want, then give instructions as exactly as you can.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2113800335919841579
This user tested how far Meta's Muse can go on video: from one character image, five 10-second clips with audio, seven made on the free plan without hitting a limit. The five clips shot on a phone were then handed to Claude Code on Opus 5.5, which inspected them in half-second steps, checked the dialogue with Whisper and used ffmpeg to assemble a side-by-side verification video against the original art. The finding was specific: close-ups, slow motion and a single spin kept the character intact, while fast movement under night neon changed the hair and clothes. The fun detail is that when asked to just make an exciting 10 seconds, Muse explained it let the surroundings carry the excitement so the character would not break.
@fabioricotta [Claude Code]
https://x.com/fabioricotta/status/2113808995890647462
The best thing this user did with Grok Bot was one instruction: install Claude Code and Codex on your own computer, hand most of the tasks to them, and work as an orchestrator that only checks quality and asks for fixes. The result was that the Grok Bot limit lasted much longer. In a follow-up the same user flagged the rough edge: when Grok Bot runs them, it is slow to notice that a task has finished, and most of the time the user has to nudge it to check.
@zsz1210 [Claude Code]
https://x.com/zsz1210/status/2113985487522447703
This user runs Claude Code and Codex on a Mac mini at home, and the Grok Bots hand work to them over Reason Network and check what comes back, so the Grok quota lasts much longer. The setup has been rebuilt many times and this version is the one that finally feels right. It is the same pattern a lot of people landed on independently this week: the hosted personal agent as front desk, the coding harnesses on owned hardware as the workshop.
@noahduck283 [Claude Code]
https://x.com/noahduck283/status/2113986208845979811
This user shared a full interconnect setup across Claude Code, Codex, Grok Bot and Gemini. Claude Code calls Codex natively and acts as foreman: because Claude reads intent so well, it drives Codex to write code and then does the quality check. Claude Code, Codex and Grok Bot talk across machines through Reason Network, and Gemini is called through a local OpenCodex endpoint for video analysis and copy polish. The notes are candid: Muse was too middle-of-the-road to bother with, and Grok Bot can connect to a lot in theory but its backend is not stable enough to be the daily driver.
@_avichawla [Claude Code]
https://x.com/_avichawla/status/2113954302901256597
An open-source project called HarnessRouter lets Grok Bot hand tasks to other agent harnesses and run them in their native runtimes, including Codex, Hermes, Claude Code, DeepSeek Harness and Pi. This user tested it by asking Grok Bot to turn a support-ticket CSV into an interactive dashboard: Grok assigned the analysis to DeepSeek Harness on Claude Sonnet 5.5, passed the output to Codex on GPT-5.5, and checked the finished dashboard in its own browser. The point is that the workers are not more Grok Bots; each delegated task runs inside a separate session with its own harness, tools and agent loop while Grok stays the coordinator. A shared protocol handles sessions, streaming, files, cancellation and failures the same way whichever harness does the work.
@longhaiqwe123 [Claude Code]
https://x.com/longhaiqwe123/status/2114031712430948410
This user is done being a human copy-paste machine between two terminals. Testing Agent Relay, an open-source tool YC just released, the user sent one sentence: Claude Code wrote the code, then called Codex to review it, the two argued in the background for two rounds, and the finished product came out the other side. The post includes a step-by-step tutorial for beginners and the full test process.
@NetworkChuck [Claude Code]
https://x.com/NetworkChuck/status/2114015265620930955
This user's favorite way to talk to agents right now is plain iMessage, with Claude Code, Hermes and Codex all set up to talk by text. The reason is practical: there is a dead zone on the morning walk that kills the terminal-app connection on the phone, and iMessage just works without a remote session. Hermes has a built-in way to connect, and for Claude Code and other terminal harnesses the user had the agent build its own bridge, which goes quickly.
@itsone_ai [Claude Code]
https://x.com/itsone_ai/status/2113841183025717617
With a pile of Claude Code and Codex windows open, the question is whether you click through them one by one or have one place to see everything. HumanLayer just open-sourced its whole core, and this user found it is the same code their cloud runs: the desktop app, the daemon that runs Claude Code and Codex sessions on your own machine, and the cloud API, with tasks, sessions, plans, comments, diffs and PRs in one interface plus GitHub, Linear and Slack hooks. It is Apache 2.0 with no telemetry when self-hosted, and the docs suggest letting a coding agent set it up for you. The pitfall the user hit: right now it will not even start without WorkOS, Slack and Linear keys filled in, which the team says is the first thing they will make optional.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2113826311512969581
Anthropic's MCP Tasks and Elicitations let one agent run long and pause to ask a human before an irreversible step. This builder made Campfire for the case where several agents are asking at once: Claude Code, Codex, Goose, Aider and OpenHands run side by side in one browser tab, and permission voting decides what executes. A majority approves an action, any single deny blocks it, and a 30-second timer stops the session from stalling on a human who stepped away. Session replay and a kanban board built from each agent's tool calls turn the approval moment into a record of who asked for what and who said yes.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2113839146028658971
A solo developer is building a Ragnarok Online meets Monster Hunter style game with Claude Code on a Max 20x plan, and shared three custom skills. Route sends easier tasks to Sonnet or Haiku with Opus reviewing; handoff starts a fresh session carrying only the context needed to continue; and Re-searcher saves research findings to Obsidian so notes do not decay into summaries of summaries. The rest of the stack is Godot, Tripo P2 for 3D assets at about 20 dollars a month, Blender, and prompts saved as JSON to keep the art style consistent. The caveat from the poster: this is relayed from the developer, and the game is still in early development.
@Tami_1001 [OpenClaw]
https://x.com/Tami_1001/status/2113999573190316346
This user once wanted to see whether OpenClaw could install a skill on its own. It failed, tried to delete the skill to reinstall it, and by accident wiped the entire 700 GB D drive, OpenClaw included. The post is a one-liner told as a joke, but it is the most concrete reminder of the day that an agent with file access and no deletion guardrail can turn a retry into a disaster.
@cupcake_trader [OpenClaw]
https://x.com/cupcake_trader/status/2113890220756398493
This user maxed out the five-hour usage window from a single day of OpenClaw use, running Sol-6.1-Low, and asked publicly what they were doing wrong. Someone from the OpenClaw side replied after looking at the numbers: 14.4 million tokens in 51 minutes, even with a 98 percent cache hit rate, and 279 tool calls, and asked for the session logs to check for unnecessary work or too much context carried between calls. It is a rare public look at where a personal agent's quota actually goes.
@ECalifornians [OpenClaw]
https://x.com/ECalifornians/status/2114127184696914194
This user gave several agents the same prompt: fill a room with everything you know about me, feel free to exaggerate. Codex, used longest, produced a crowded room that covered AI work, a Mac mini fleet, camping, family, photography and faith, which the user found a fairly rounded portrait. Grok Bot only knew the user's X activity, which is mostly election campaigning, so its room was a campaign machine and nothing else; Muse, the newest tool, produced a tidy room because many projects had been explicitly removed from it. OpenClaw, the first tool the user ever used heavily, was not even asked, because it has barely been touched lately.
@utekkare [OpenClaw]
https://x.com/utekkare/status/2113788551620087878
Over a weekend this user wired Cursor locally through OpenClaw and ran Codex through a custom Slack bot called Codex Review to build two new satellite apps for their company. One is a customer service tool for brands; the other is a commerce satellite that runs Shopify, wholesale and Amazon from one integrated warehouse system and product feed. The sprints coordinate through a shared Slack channel, Linear tickets and GitHub PR comments, so the agents work in the same places a human team would.
@Jordan_RK [OpenClaw]
https://x.com/Jordan_RK/status/2114071366014284119
This user does not want to spend Grok Bot credits on narrow, repeatable chores, so each one gets its own cheap lane. Adding a film to Jellyfin goes to a Hermes agent topic in Telegram: drop the file path and it loads everything. OpenClaw handles narrow tasks fired from n8n calls, where a Telegram topic, a skill and local files reachable from one API call work well. Everything else goes to a CLI or Grok Bot, which is a tidy example of routing by cost rather than loyalty to one agent.
@PavelVanchugov [OpenClaw]
https://x.com/PavelVanchugov/status/2114009802288320539
This user runs OpenClaw on a standalone Dell server as a backup, reachable over Tailscale. The Grok Bot also uploads its own memory to a shared brain on the same server, so OpenClaw and Grok Bot share memory and information while staying separate systems. The Grok Bot also borrows OpenClaw's multi-subscription inference to save its own tokens, and the result is surviving a week of fairly heavy use on a 60-dollar subscription.
@kyle_e_walker [OpenClaw]
https://x.com/kyle_e_walker/status/2114078012484284811
This user prefers OpenClaw on a VPS because of a Slack-based setup and full control over which models run. On top of that, a local Opus agent can SSH into the VPS and handle maintenance just fine. The user's read on the bigger debate is fair: the steady stream of people switching to hosted agents is the real counter-argument to the idea that everyone will roll their own software.
@designertom [OpenClaw]
https://x.com/designertom/status/2113814938888380823
After more than ten months on OpenClaw, this user is slowly migrating. For any subject that needs an always-on expert, where OpenClaw subagents would once have been spun up, the user now creates a Grok Bot instead. The high-end work that needs more customization, recall and memory stays on OpenClaw, mainly because it allows control over the model and thinking level, and because trust in Grok Bot is still being built case by case.
@srechakrareddy [OpenClaw]
https://x.com/srechakrareddy/status/2114080746319282594
This user originally ran OpenClaw or Hermes to own the memory map and lists and be able to carry them to other assistants, but the constant maintenance became too much. The fix is a hybrid: the memory map, to-do lists and everything else are kept separately, and any or all of the agents update them. Whether the work happens in Grok Bot, Muse or Duet, the memories behave like a service that every agent reads from, which keeps the original reason for self-hosting without the upkeep.
@helloitschrisg [OpenClaw]
https://x.com/helloitschrisg/status/2113990881933918433
This user has run OpenClaw and Hermes since the start and still tinkers with Hermes, but agrees with the complaint that the maintenance makes it a hobby. Grok Bot now does the actual work in the user's businesses. If Grok Bot and Dots keep improving at this pace, the user says there is no real need for the rest. It is one voice, but the same sentence showed up from many long-time self-hosters this week.
@drmhse [Claude Code]
https://x.com/drmhse/status/2114051274245017852
The first thing this user does on any new machine is tell Claude Code to memorize globally to never use subagents. The reason is measured from experience: subagents drain usage more than twice as fast, because each one takes in the full context no matter how narrow its task is. It lines up neatly with Anthropic's own advice this week to fold unnecessary subagents back into one loop.
@TsuruRonin [Claude Code]
https://x.com/TsuruRonin/status/2114087217748455846
This user runs a tower crane and stops lifting once the wind hits 15 meters per second. Anthropic's new guidance says subagents should be judged by a number too: build the eval first, then fold them in one at a time and watch the score. The user checked and found subagents took 55 percent of their biggest Claude Code session, with no eval in place yet. The crane analogy says it well: a hard threshold is only useful if you are measuring.
@YoussefHosni951 [Claude Code]
https://x.com/YoussefHosni951/status/2113814091185733766
This user usually starts a new Claude Code or Codex session at 50 percent context, but that rule is not always reliable. The better signal is a trick credited to Peter Steinberger, OpenClaw's creator: one line in CLAUDE.md that tells Claude to start every response with your name. It works as a canary, because the model tends to drop small rules before it starts inventing things, so when your name stops appearing it is time for a fresh session.
@sunsetsyntax [Claude Code]
https://x.com/sunsetsyntax/status/2113801560111415354
This user hit Claude Code's rate limit in the middle of a refactor and lost the thread again. Sno Station fixed it: Codex and Claude Code share one local encrypted memory, so Codex carried on with the context intact, with no account and no daemon. The one catch the user flags is to switch off the nightly session upload in the settings file.
@zhu185178 [Claude Code]
https://x.com/zhu185178/status/2114061982014021991
This user has uninstalled almost every skill, both in Codex and the ones written for Claude Code, because current models do the tasks as well or better without them, especially for video. The observation is sharp: when a great one-shot product video gets saved as a skill, reusing that skill produces worse videos than the original free run, because a skill is both a floor and a ceiling. The user now writes skills only for SOP-style work that must reproduce exactly, and for beginners who cannot yet describe what they want, where the skill really guides the person rather than the agent.
@aryanXmahajan [Claude Code]
https://x.com/aryanXmahajan/status/2114061811658215727
This user runs one research loop on a Mac every night. At 2 a.m. background Apify actors pull the top 50 viral posts in the niche from X and TikTok, new ads from Meta Ad Library and Reddit threads where buyers complain about competitor tools; Firecrawl or Jina turns them into clean markdown saved as local files. Then Claude Code in the terminal analyzes hooks, angles and the exact phrases buyers use, flags angles everyone is repeating that stopped working, and spots questions nobody is answering. By morning the local brief has three fresh content angles backed by real view counts and five competitor weaknesses to attack.
@geraldang711 [Claude Code]
https://x.com/geraldang711/status/2113792497076011374
As the first go-to-market hire at a VC-backed startup, this user lists the boring ways AI actually makes outbound better: research a company in under a minute, find the reason to email today, write only the first line, and tag replies so only interested and not-now get read. The Claude Code part is the most durable win: a scraper that broke every week was replaced with a small tool built in-house that just works. The broader point is that AI does not do outbound for you, it makes the parts that matter faster.
@kinryuchan [Claude Code]
https://x.com/kinryuchan/status/2114103476930441607
This post relays a demo by the Japanese entrepreneur Takafumi Horie, who built an app with Claude Code in about an hour that a contractor would quote in the millions of yen. It is a four-choice quiz app built from 17 rounds of past exam questions posted on a Japanese ministry website. The whole instruction was sending the URL and saying make an app, and the only bug in the first version was that figures and tables did not display.
@murasametech [Claude Code]
https://x.com/murasametech/status/2113828012072554959
For home security work this user split roles between cloud AI and a local model. Claude Code in the cloud checks whether news is true, judges impact, designs the procedure and verifies results; a local Qwen model on a home GPU does the hands-on steps; and the human makes decisions and runs anything that needs sudo. The hidden benefit is that dangerous material never leaves the house: server login keys, raw command output and config contents stay local, and the cloud model only sees summaries and the work report. The other lesson was to state up front that this is protecting your own equipment under your own approval, because with a clear purpose and scope the cloud model helps without hesitating.
@beetz12 [OpenClaw]
https://x.com/beetz12/status/2113843876817416259
A consultant pitched a full multi-agent system to a Shopify wellness founder and was turned down, and thinks the founder was right. Six months ago an agent meant fighting frameworks like OpenClaw and Hermes; this founder had already built a company brain from brand docs, connected support and ads tools and wired Claude into the email platform. The founder only kept the door open for operations and inventory math, which matches the consultant's line: drafting work with a human reviewing can be built in-house, while the last 20 percent, where agents touch live customer data, bank feeds and production keys, is where a fragile setup does real damage.
@junjun_math [Claude Code]
https://x.com/junjun_math/status/2113836915875283075
This user debugged Claude Code's remote control from the logs. When a new session created on the PC does not show up on the phone, send the first message from the PC and it appears, because remote control cannot attach to a session before any exchange has happened. Right after waking from sleep it will not connect for about three minutes, and the write-up lists each cause in the order they were confirmed.
@avive [Claude Code]
https://x.com/avive/status/2113996164890611921
This user spends 10 to 16 hours a day with Claude Code and Codex, and by noon the walls of text were wearing them out. The answer was action cards, a compact format that turns each agent update into something scannable instead of paragraphs. It is a small interface fix, but it targets a real cost that nobody budgets for: the human reading time on long agent sessions.
@nicolaycz [Claude Code]
https://x.com/nicolaycz/status/2114082206637580717
This creator runs a personal brand with a team of Grok Bots, one for each social network's content, one for leads, one for the platform and one for the inbox. The videos are edited with Claude Code inside Iris, an editor the creator built. What matters most in this setup is a coordinator plus specialists that each work with their own tool, which is why the idea of Grok Bot coordinating and passing tasks to Claude Code or Codex caught the creator's attention.
@gdmvx1 [Claude Code]
https://x.com/gdmvx1/status/2113993361858195896
This user has been doing a manual version of agent routing for a while: Grok Bot runs the user's gym with about 20 bots and hands the heavy build work to Claude Code. A router layer like the newly open-sourced HarnessRouter would replace a lot of hand-written glue. It is a short post, but a gym run by a bot team that outsources coding to Claude Code is exactly the non-developer shape this feed keeps finding.
@ChrisGPT [Claude Code]
https://x.com/ChrisGPT/status/2113796109722153067
This user had Kimi K3.1 Pro make a trailer for Claude Code and compared it with Opus 5.5. The visuals were surprisingly a bit better than what Opus 5.5 produced, although it reused many of the same stock-sounding effects that frontier models keep picking. The verdict is split: Kimi delivers on visual work, but for general-purpose and white-collar tasks Opus 5.5 is still more useful, and a better-looking trailer does not make a better model across the board.
@shuntnoce [Claude Code]
https://x.com/shuntnoce/status/2113955674871013633
This user ran the same prompt through one harness on three models, all at high effort and with no edits: make a single slide explaining how hemodialysis works. The outputs came from Claude Code on Opus 5.5, Codex on GPT-6.1 Sol and Gemini 3.8 Flash. The judgment was practical: Claude and Gemini made the more presentation-ready slides, while the GPT one read better for actually delivering the talk.
@ReignSwords [Claude Code]
https://x.com/ReignSwords/status/2114092130809651295
This user reviewed GenOffice, an open-source office suite trending on GitHub, without opening a single app, because the CLI is what the agents actually use. Markdown to .docx took 3.8 seconds and opened cleanly, JSON to .xlsx took 0.9 seconds with a live SUM formula returning exactly 42, and .docx to PDF failed twice until a virtual display was provided. The part that matters is byte-preserving edits, so only what changes gets rewritten and the file still opens clean in Word, plus a skill file so Claude Code or Codex can build a real .pptx headlessly. The cold water: the 188 MB download filled a 512 MB /tmp and killed the first install, and letting AI edit real work files still needs tracked changes and one-click rollback that work every time.
@LeNoobCodeur [Claude Code]
https://x.com/LeNoobCodeur/status/2113900132160635197
This user points out the open question in MCP Tasks. A compatible server can keep a job running after you close Claude Code, and Claude can later find its state and fetch the result, which is handy for jobs that take minutes. But if the server loses track of the job, there is no way to know whether the action already happened, and relaunching it risks running it twice. For anything that sends, pays or deletes, idempotency is the part the protocol still leaves to you.
🗣 User Voice
User Voice
Quota is still the loudest pain, and now it is harness-shaped. @DeepakNesss burned a whole five-hour Max window in about 30 minutes through T3 Code with the same prompts used in the CLI; @cagrimbakirci reports two to three times the usual Opus 5.5 burn for several days; @Tony_Stef_ runs four Claude Code accounts next to two Codex accounts and says Codex resets faster than it can be used up while Claude Code never does. @drmhse bans subagents on every new machine because each one swallows the full context.
Self-hosting fatigue is real, but control is what people refuse to give up. @mikekarnj asks whether anyone still maintains OpenClaw when Grok Bot works out of the box, and @tagomago_me is staying but calls maintenance the cost of sovereignty. @bmagierski finds Anthropic models through OpenClaw expensive because they force the API instead of a Teams plan.
People want the handoff between agents to be cheap, and doubt it always pays. @KTrifonov has tried several ways to pass tasks from Grok Bot to Codex or Claude Code and has not seen a real efficiency gain yet; @fabioricotta says Grok Bot is slow to notice that a delegated job finished; @DavidJonR found max reasoning trips a Hermes timeout, so long-thinking work has to live elsewhere.
Permissions need to travel with the work. @mano_sec asks whether a router keeps each harness's own permission model, since the same task under a Codex sandbox and a Claude Code allowlist touches different things; @AmrTawfik160 notes most Claude Code setups on GitHub have zero deny rules; @LeNoobCodeur wants to know how a background MCP task avoids running twice when the server loses track of it.
Users still rank Claude Code highest on reading intent. @JoshExile82 says Claude Code understands user intent and context far better than Grok Bot, which kept taking away one example from an hour-long conversation; @jatinsj710 wants one install path for a skill across Claude Code, Copilot and Codex instead of an adapter for each.
Quota is still the loudest pain, and now it is harness-shaped. @DeepakNesss burned a whole five-hour Max window in about 30 minutes through T3 Code with the same prompts used in the CLI; @cagrimbakirci reports two to three times the usual Opus 5.5 burn for several days; @Tony_Stef_ runs four Claude Code accounts next to two Codex accounts and says Codex resets faster than it can be used up while Claude Code never does. @drmhse bans subagents on every new machine because each one swallows the full context.
Self-hosting fatigue is real, but control is what people refuse to give up. @mikekarnj asks whether anyone still maintains OpenClaw when Grok Bot works out of the box, and @tagomago_me is staying but calls maintenance the cost of sovereignty. @bmagierski finds Anthropic models through OpenClaw expensive because they force the API instead of a Teams plan.
People want the handoff between agents to be cheap, and doubt it always pays. @KTrifonov has tried several ways to pass tasks from Grok Bot to Codex or Claude Code and has not seen a real efficiency gain yet; @fabioricotta says Grok Bot is slow to notice that a delegated job finished; @DavidJonR found max reasoning trips a Hermes timeout, so long-thinking work has to live elsewhere.
Permissions need to travel with the work. @mano_sec asks whether a router keeps each harness's own permission model, since the same task under a Codex sandbox and a Claude Code allowlist touches different things; @AmrTawfik160 notes most Claude Code setups on GitHub have zero deny rules; @LeNoobCodeur wants to know how a background MCP task avoids running twice when the server loses track of it.
Users still rank Claude Code highest on reading intent. @JoshExile82 says Claude Code understands user intent and context far better than Grok Bot, which kept taking away one example from an hour-long conversation; @jatinsj710 wants one install path for a skill across Claude Code, Copilot and Codex instead of an adapter for each.
📡 Eco Products Radar
Eco Products Radar
Codex: the second harness in almost every stack, as reviewer for game payment code, as co-prover on Rule 30, and as the comparison point for limits and permission prompts.
Grok Bot: the hosted personal agent people now use as dispatcher, handing builds to Claude Code and Codex on their own machines.
Hermes: the self-hosted agent most often named as where OpenClaw users went, and one of the harnesses behind iMessage and Telegram setups.
Muse, Dots and Claude Duet: the hosted personal agents users compare on memory, video and proactivity.
Cursor: the third harness in multi-agent setups, wired through OpenClaw and Linear.
Opus 5.5: the model behind most of the day's builds, and the one users say is burning quota faster this week.
Kimi K3.1 Pro: the new open-weights model people tried inside Claude Code and against Opus 5.5.
DeepSeek: the cheap model and harness routed in for analysis work.
MCP Tasks and Elicitations: the long-running job support that people measured, explained and questioned on idempotency.
HumanLayer: the agent control plane that was open-sourced with its desktop app, daemon and cloud API.
HarnessRouter: the open-source router that lets Grok Bot run tasks inside Claude Code, Codex, Hermes and Pi.
Pi: the minimal harness that shows up in routers and control planes next to Claude Code and Codex.
Codex: the second harness in almost every stack, as reviewer for game payment code, as co-prover on Rule 30, and as the comparison point for limits and permission prompts.
Grok Bot: the hosted personal agent people now use as dispatcher, handing builds to Claude Code and Codex on their own machines.
Hermes: the self-hosted agent most often named as where OpenClaw users went, and one of the harnesses behind iMessage and Telegram setups.
Muse, Dots and Claude Duet: the hosted personal agents users compare on memory, video and proactivity.
Cursor: the third harness in multi-agent setups, wired through OpenClaw and Linear.
Opus 5.5: the model behind most of the day's builds, and the one users say is burning quota faster this week.
Kimi K3.1 Pro: the new open-weights model people tried inside Claude Code and against Opus 5.5.
DeepSeek: the cheap model and harness routed in for analysis work.
MCP Tasks and Elicitations: the long-running job support that people measured, explained and questioned on idempotency.
HumanLayer: the agent control plane that was open-sourced with its desktop app, daemon and cloud API.
HarnessRouter: the open-source router that lets Grok Bot run tasks inside Claude Code, Codex, Hermes and Pi.
Pi: the minimal harness that shows up in routers and control planes next to Claude Code and Codex.
Comments