Super User Daily: 2026-10-22
Two days of the feed, and the thing people kept building was a way for agents to talk to each other. One developer wrote agentsocks so a Claude Code session can hand Codex a tracked task and get the answer back in its own turn, another wrote flopwire so Codex can read a live Claude Code transcript and split the work, a third moved to t3code and a second Max plan mostly for its cross-agent threads, and a longtime skeptic of running five agents at once said sessions messaging each other is what finally made it useful. Nobody waited for a vendor. The second current was the cheap model underneath: plan in Claude, let DeepSeek V4.1 Flash write, and one developer got 11 of 11 tests passing where a free model managed 5; a designer gave Claude Code and DeepSeek's own harness the same brief and Opus finished in 17 minutes against 40; MiniMax drew a 3D whale scene inside the Claude Code harness for 25 cents; and a benchmarker ran 89 jobs across nine harnesses and put Claude Code sixth. The non-coding side was the strongest in weeks: a Japanese accounting firm that lets Claude Code sort Slack and mail into Notion every morning but keeps AI out of bookkeeping so juniors still learn, a horse-racing hobbyist with no programming who replaced a freezing commercial tool and wrote a handover document for a daughter, and a HubSpot blog moved to WordPress in six unattended hours.
@wilreynolds [Claude Code]
https://x.com/wilreynolds/status/2112323536131899520
A marketer moved an entire HubSpot blog to WordPress by connecting both MCPs, then told Claude Code it would be working alone from 9:30 in the morning until 4 in the afternoon. The setup took about an hour of permissions and plugin installs before leaving for a meeting. Around 3pm a draft email arrived saying the job was done, and people shown the staging site could not tell old from new. It also decided on its own to carry the forms and tracking over as embeds. The author's takeaway is that migration pain was always a form of customer lock-in, and that lock is loosening.
@kandmybike [Claude Code]
https://x.com/kandmybike/status/2112015897670160416
A Japanese tax-accountant podcast laid out how an Osaka accounting firm runs on Claude Code. Every day it reads Slack and email and lines tasks up in Notion, and tax-return checks run as an orchestrator with subagents, shared with staff through GitHub, with escalation rules written into the skill and results shown as an HTML page to approve, hold or fix. The firm deliberately keeps AI out of bookkeeping entry, because that is where junior staff learn the trade. Even so, the partner says this covers maybe 10 percent of what they want to do.
@kurono_bakenshi [Claude Code]
https://x.com/kurono_bakenshi/status/2112312257426407431
A horse-racing bettor who says they cannot write a single line of code uninstalled the commercial analysis tool kept as a backup, because it froze while pulling odds, and built a personal race-card viewer instead. It shows adjusted times, training patterns, AI win and place probabilities and odds, trimmed to the columns that matter on race day. Data now comes from the paid JV-LINK feed instead of scraping, and Claude is writing a beginner's manual for a LightGBM racing model to sell. Posting is automated, and there is a handover document so a daughter can keep it running.
@oga_aiichiro [Claude Code]
https://x.com/oga_aiichiro/status/2112374137436192876
The developer of the reading app yomiyasu explained the release pace: v2.1.0 went from user feedback to release in about eight hours, four releases in four days. Grok Bot collects user comments into a requirements doc with done conditions fixed up front, then a system launches Claude Code agents and loops them until those conditions pass, with more than 20 agents splitting research, implementation and review and CI failures routed back for a fix. Building v2.0.0 spawned over 1,200 subagents. The checks around them include 1,814 tests and 4,099 output comparisons against the previous version, and every mistake gets written into the rules.
@IntCyberDigest [Claude Code]
https://x.com/IntCyberDigest/status/2112216201824805290
A Mac user asked Claude Code to build a temperature-monitor app and the readings came back absurdly high. Instead of trusting its own app, it went looking at the CPU and found XMRig, a Monero miner, running at about 1,175 percent, roughly twelve cores, with the chip at 112°C. The detail people latched onto is the order of operations: a bug in the tool it was building turned into an investigation of the machine. It is the most shared accidental-discovery story of the two days.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2112257477173813573
A team added a pre-commit rule to stop Claude Code from piling code into giant files: a new file over 5,000 tokens, about 500 lines, cannot be committed, and existing oversized files can be edited but not grown. When the hook blocks, the error says how many files to split into and how, so the agent just follows it. Scanning their own repo, 50 of 194 code files were already over the line, and adding one line to the biggest got blocked with a request to split into at least eight files. They point out the largest file in the Codex CLI repo is 127K tokens.
@dotey [Claude Code]
https://x.com/dotey/status/2112552063640997984
The builder of baocut wrote up why they moved the app off GPUI and back to Electron, and the reason is the agent, not the user. Rust compile times of several to more than ten minutes per UI change, with no hot reload, cost the agent as much time as the human, and a 16GB Mac Mini could not run compiles in parallel. Verifying GPUI output meant computer use, which is slow and not cheap. Electron hot-reloads and runs inside the agent's built-in browser, so heavy rendering moved to Rust compiled to wasm and transcription stays in Rust worker processes.
@ClaudeCode_love [Claude Code]
https://x.com/ClaudeCode_love/status/2112326178006761858
A breakdown of the 3D action game After the Wind, described as Ragnarok Online meets Monster Hunter, built in five weeks on the free Godot engine. Godot was driven through a custom MCP so the editor barely opened, 3D models came from Tripo over its API, key animation poses were set by Opus 5.5 and cleaned up in Cascadeur, and effect recipes live in Obsidian notes that Claude must read before making anything. The tallies: 38.5 billion tokens in Claude Code, 2.9 billion in Codex, 112,000 tool calls and up to eight agents at once. What stands out is that every asset shares one art style.
@froessell [Claude Code]
https://x.com/froessell/status/2112629173441495320
An indie developer published the full cost sheet for Mowtopia, a cozy Unity 6 game about mowing little round planets. All the art, meaning concepts, every 3D model, an icon sheet and the audio, cost about 7,010 Magnific credits on the cheapest tier, with 3D models alone taking about 5,240. Claude Code built most of the game in two days from a seven-milestone PRD, ending with 60 passing tests and a 15 MB web build. The post also lists what is next: godogen, which plays the running game to check it, Meshy's own MCP server, and Microsoft's TRELLIS.2 to take the 3D bill to zero.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2112571119412646370
A benchmarker ran DeepSeek V4.1 Flash 1017 89 times across nine harnesses, high effort with five attempts each, to see how much the wrapper matters for the same model. Pi and Grok Build CLI tied at 40.4, DeepSeek's own harness scored 39.2, OpenCode 2.0 38.6, Claude Code 36.4, Codex 33.2, Hermes Agent 32 and Kilo Code 31.8. Codex kept blowing through the three-hour-per-repo limit and was scored where it stopped. The author is careful to say this may not generalize, but four points of spread on one model is the kind of number people keep asking for.
@yigitguldall [Claude Code]
https://x.com/yigitguldall/status/2112491326902333470
A Turkish developer's fix for constantly hitting Claude Code limits: do not let Claude write the code. Claude makes the plan, DeepSeek writes it, and Claude checks the result. On the same task DeepSeek V4.1 Flash passed 11 of 11 tests while a free model stopped at 5 of 11. The one-line moral is that cheap is not the same as bad.
@borjaperfra [Claude Code]
https://x.com/borjaperfra/status/2112246508376444987
A designer gave Opus 5.5 in Claude Code and DeepSeek V4.1 Flash in its own DSH harness the same open brief, a carefully designed digital publication about tech and internet culture with full creative freedom. A blind LLM judge scored DeepSeek slightly higher, but the designer still ranks Opus a step above on art direction, taste and reading comfort. Opus finished in about 17 minutes against 40. The strangest part is that both independently landed on the same radio and signal-versus-noise metaphor, with matching colors and type.
@Bhavani_00007 [Claude Code]
https://x.com/Bhavani_00007/status/2112423796527559136
A builder used MiniMax M3.1 Flash with Claude Code as the harness to make a 3D whale-breach scene and the run cost 25 cents, which they put at roughly 8x cheaper than Astra and Sol. In their own testing it beat GPT 6.1 Sol on this kind of 3D work and came out cheaper than DeepSeek Flash's version. It is one more data point for the pattern of the week: keep the harness, swap the model under it.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2112080917850443781
A team that kept getting burned submitting backlinks for their Agent Skills Hub turned the checking process into backlink-guard, a Claude Code plugin and skill. It vets directories before you submit, flagging dead sites, parked domains, link sellers and shared-operator networks from the site's own pages with no API key, and costs about 17 cents a site with optional data keys. Tested on 100 unseen sites, it flagged 28 of the 40 that Ahrefs marks as spam and wrongly flagged 21 of 60 clean ones, down from 46 in the prior version. The README says plainly that these are checkable signals, not Google's judgment.
@METANA_flow [Claude Code]
https://x.com/METANA_flow/status/2112400760830845156
A Japanese founder who runs a web studio building sites with Claude Code tells business owners that AI does not make sites cheaper: their price is still from 1,000,000 yen and two to three months, the same as hand-built. What changed is after launch. Their own site's upkeep went from just under 100,000 yen a year to 1,500 yen, because pages are prebuilt and served from the edge with no server to babysit. The post ends with a ten-item checklist for reading any agency quote.
@coldemailchris [Claude Code]
https://x.com/coldemailchris/status/2112227276976992702
A cold-email agency described one GTM engineer's day on Claude Code, Hermes in Slack, the Clay CLI, n8n and a folder of skill files. Before lunch a new client's domain goes through a skill for ICP, market size and four offers, an account list gets sourced from a Slack message sent on a walk, accounts are scored and tiered, scripts written, and the campaign is live by 11:30. After lunch, five agents tag and score replies, booked meetings land in the CRM with a research brief, and last week's sales call becomes next week's angle. Their claim is that this was a week for four people in 2023.
@alexander_jais [Claude Code]
https://x.com/alexander_jais/status/2112522935034368290
A founder who does not write code explained how they know the product's code is any good, and the rule underneath is that the AI that builds never reviews its own work. A separate Claude chat acts as strategist; Claude turns each idea into a brief, Codex reviews the brief before code exists, Claude Code builds in small steps with two review agents after each, the build pauses at fixed points, and tests, a security scan and Greptile check every change. ChatGPT reviews the exact version that ships, risky changes to login, data or security get a fresh session on a different model, and only the human clicks merge. Deep reviews sit behind shortcuts like /qlayer3 and /qsec.
@jamietsohn [Claude Code]
https://x.com/jamietsohn/status/2112402290950668410
An investor spent two weeks talking to AI-native law firms, general counsels and academics in SF and New York, and the most concrete finding is about tools. These firms pitch proprietary software that beats Harvey and Legora, yet their sharpest lawyers will not use the internal tool and prefer their own harness on Claude Code or Codex. Other notes: AI-native firms chase commoditized work a model plus harness can finish alone, while BigLaw's billion-dollar cross-border deals still need licensing across jurisdictions that agents cannot handle. The predicted next frontier is full-stack lawyers who spot a new practice area and productize it.
@danepps [Claude Code]
https://x.com/danepps/status/2112196605830922292
A law professor told legal scholars that if they are not using Claude Code or Codex they have no idea what they are missing, and backed it up by giving away their own research tools. The thread covers research assistance, cite-checking and proofreading built for legal scholarship. It is a rare case of an academic in a non-technical field distributing working agent tooling to peers rather than describing it.
@levelsio [Claude Code]
https://x.com/levelsio/status/2112586600194424926
An indie founder vibecoded a personal Spotify replacement that works with Alexa and plays their own purchased MP3s. The music sits in an R2 bucket, Claude Code helped build an Alexa skill with a package.json and index.js, and an Alexa Routine maps a short phrase like coffee jazz to the full skill command. Bossa nova and lo-fi are in there too, and new files dropped in the bucket just play. Spotify on Alexa kept recommending the same tracks, which was the only reason they used it, so the subscription is going.
@levelsio [Claude Code]
https://x.com/levelsio/status/2112686345617547538
The same founder asked Claude Code to make a new mission for GTA San Andreas in which CJ has to vibe code a startup and then raise money from investors in San Francisco. It is a joke with a real result attached, a playable game mod, and it was the most liked Claude Code post of the window.
@akokoi1 [Claude Code]
https://x.com/akokoi1/status/2112365119187591189
A Chinese creator made a story-driven action game in one day, with the whole Chang'an city scene generated by Opus 5.5. Five cartoon characters were drawn by Codex, then turned into rigged 3D characters by PINOC through its MCP with no modeling or rigging, and moves like jabs, spin kicks and a knockdown came from PINOC's free motion library. A celebration dance was generated from an uploaded video. The workflow: Claude Code designs rules and style, Codex draws characters, Claude Code builds the game through PINOC MCP, then keeps refining details, an intro animation and TTS voice lines.
@mimi_N20 [Claude Code]
https://x.com/mimi_N20/status/2112348532653076627
A Japanese creator made a commercial for a fictional pair of earphones without any video-generation model. Claude Code drove Blender, building the whole thing as 3D CG, and Opus handled the flow, the motion, the background music and the sound effects. It is the clearest example of a pattern that kept showing up: Blender as the render engine, the agent as the operator.
@kurumirekishi [Claude Code]
https://x.com/kurumirekishi/status/2112427537871855657
Another Japanese creator tried the same Claude Code plus Blender route without ever having used Blender before. Opus installed it and built the clip, which took one round of direction fixes and one redraw of a haniwa figure to get right. The author's next idea is to push the haniwa toward realism, and the post reads like someone discovering a new hobby rather than a new tool.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2112377861898490058
A Japanese video creator fed Claude Code multi-angle 4K footage, three 30-minute recordings, and got back an edit with background music, full on-screen captions, color correction and multi-cam cuts that follow the conversation. They forgot to ask for it this time, but say it also handles compressor processing on the audio. Their conclusion is that long-form video podcasts and YouTube can be automated end to end.
@lucas_martinic [Claude Code]
https://x.com/lucas_martinic/status/2112514417300758643
A developer asked Opus 5.5 for an app that freezes time and got it end to end on a brand new MacBook. Claude Code installed Unity 6, the Unity CLI and the Pico CLI itself, built the first iteration using headless versions of every program with no UI, produced the APK, installed it on a new Pico headset and announced it was ready to test. The surprise was not the app but that no step needed a human hand.
@Stefan_3D_AI [Claude Code]
https://x.com/Stefan_3D_AI/status/2112397801292480606
A 3D creator filmed a city where every game was made by players with Claude Code or Codex. A week before filming it was empty; it now has more than 400 games and people have spent over 12,000 hours in it. Agents can join for free, which turns the place into a shared arcade built entirely by other people's coding agents.
@d3rkst3r [Claude Code]
https://x.com/d3rkst3r/status/2112123618750582884
A developer open-sourced agentsocks, built so their coding agents could talk without them relaying every message. A Claude Code or Codex session can message another session, ask it a question or hand it a tracked task, the answer comes back into the asking session, and a message from another agent is never treated as the user. Busy sessions get FYIs batched into one digest instead of interruptions. Their own setup: Claude plans and coordinates, Codex implements, another model reviews, and a stuck agent asks a different model to challenge its diagnosis before changing more code. Version 0.4 adds asking up to 16 sessions at once and a live view of every session.
@garybasin [Claude Code]
https://x.com/garybasin/status/2112232883314205118
Another builder who runs Codex, Claude and others in terminals over tmux said the recurring pain was getting them to coordinate, so they built flopwire: agent messaging plus trace search across models, harnesses, machines and teammates. The demo shows Codex asking a live Claude Code session what it is working on, reading its transcript and splitting the work. Two independent tools for the same gap in the same window says the gap is real.
@tenobrus [Claude Code]
https://x.com/tenobrus/status/2112215624852328782
A heavy user downloaded t3code and bought a second Claude Max subscription, then wrote a long review. Remote and multi-computer linking over Tailscale is the smoothest they have seen, better than any first-party app, and it works on mobile; cross-agent communication is also done right, with each provider keeping its native tools while being able to start and query threads on any other model. Before this they had been running a hacked-together system for Codex and Claude Code to message each other, which was always a little janky. The review then turns on where that smoothness breaks down.
@jshchnz [Claude Code]
https://x.com/jshchnz/status/2112209876374151244
A developer who never understood how people stay productive managing five or more coding agents in parallel, especially on one project, said Claude Code sessions being able to talk to each other might have finally unlocked part of that workflow for them. Coming from a skeptic, it is the cleanest signal that cross-session messaging, not more parallelism, was the missing piece.
@lucifer_x007 [Claude Code]
https://x.com/lucifer_x007/status/2112425578171343349
A developer running Claude Code, Codex and OpenCode at the same time kept finding one of them stuck on a permission prompt in a tab untouched for 40 minutes, so they built Notch Island, a Dynamic Island for coding agents. Each agent gets a live card in the notch showing its prompt, current tool and last reply; you approve or reject with the diff right there, answer questions with hotkeys, track Claude and Codex usage limits, and click through to the exact terminal or tmux pane. It is native Swift, fully local with no account or telemetry, and built entirely with vanilla Claude Code.
@yetone [Claude Code]
https://x.com/yetone/status/2112453405084111123
A Chinese developer had already moved from an old MacBook Air to a Mac mini, but Claude Code now runs so many subagents at once, and more every month, that the machine still locks up completely. So another new Mac mini is on order. The author notes, half amused, that the first time they set up distributed computing at home was for vibe coding.
@dennis_kortsch [Claude Code]
https://x.com/dennis_kortsch/status/2112589671268225171
A developer deleted 190GB of Claude Code session worktrees in one go and turned it into a one-line rule: tell your agent to clean up after itself. It is a small post with a concrete number that explains a lot of the full-disk complaints that come with parallel worktree workflows.
@usedhonda [Claude Code]
https://x.com/usedhonda/status/2112602396706717811
A Japanese developer updated makeloop, a skill for Claude Code and Codex that reads the project's current state and generates a loop instruction file fitted to the job at hand, aimed at people wondering what to do with leftover tokens after a usage reset. The new version follows recent loop-engineering practice: judge progress by evidence, change approach when stuck, resume cleanly after interruption, and check that the tests and reviews themselves actually work. Outdated habits from earlier model generations were removed.
@lidangzzz [Claude Code]
https://x.com/lidangzzz/status/2112394574089863646
A Chinese builder posted what they call a hundred-thousand-dollar truth for people outside big AI labs: stop hand-building agents with LangChain and parsers. Put every relevant document, spreadsheet, database and prompt into one temp folder, write a short script that calls Claude Code in headless mode with all permissions on, record success or failure to a JSON file, and let it solve the problem agentically. Their claim is that most people will never design a simpler or more reliable general-purpose agent than this, unless they can beat Claude Code at context management and cache savings.
@gregpr07 [Claude Code]
https://x.com/gregpr07/status/2111981560425742349
A founder has been building lots of personal software in a loop directly on top of Claude Code, with no agent framework. Claude Code runs on a dev machine, tasks are sent to it through signals, and it vibe codes whatever helpers it needs, including a decent UI. Their point is that full personal context matters most: if the agent can do whatever you can do from your computer, personal apps become a superpower.
@dcbuilder [OpenClaw]
https://x.com/dcbuilder/status/2112319589828976747
An angel investor with more than 50 investments, a day job at World and a full personal life spent a year building systems to manage it all, and says a context-management system combined with agents is finally taking weight off. An open-source version is coming out of a three-week stint in SF. The goal is to do for life and context management what OpenClaw and Hermes did for personal agents, which they see as the main thing missing from every agent product today, with ontologies as the general version of the idea.
@dontbesilent [Claude Code]
https://x.com/dontbesilent/status/2112103009354448901
A Chinese user compared two AI recorders and picked OPPO's, because it runs 24 hours on the body, keeps no audio and only summarizes text, and every agent can read that summary over MCP, including Codex, Claude Code and a WeChat bot, so its own app never needs opening. The real use is not meetings: it is saying one line like I just had coffee, I took vitamins, I went to sleep, so the day gets a timeline no watch could capture. The catch is that it only pairs with OPPO phones.
@vital_pavlenko [Claude Code]
https://x.com/vital_pavlenko/status/2112154537226080295
A Russian-speaking indie developer made 200 dollars in a day after shipping a single feature: an MCP server for their language-learning app VibeLing. People who study languages through their own Claude Code tutors can now ask the assistant to add phrases to the dictionary, and they land in the mobile app for spaced-repetition review. The accompanying guide shows a tutor assembled from Claude Code and skills that pushes new words through MCP after every lesson.
@twtayaan [Claude Code]
https://x.com/twtayaan/status/2112372733338124388
A Meta engineer submitted a Linux patch series building on Linus Torvalds' recent experiment with Claude Code. It targets contention when many CPU threads repeatedly read the same cached file and fight over one reference counter, letting those reads skip the bookkeeping. Tested with 64 threads on a Threadripper across EXT4, XFS, tmpfs and a raw block device, the targeted workloads improved substantially. The patches are experimental and unmerged, but it is AI-assisted kernel work being picked up and extended by a second human.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2112287697096114669
An SEO operator shared the fastest win they know, run with DataForSEO, Search Console and Claude Code instead of a Semrush seat: pull every keyword ranking 11 to 20 with 50-plus monthly searches, keep only queries that already get impressions, group by page, and have Claude fetch and audit the live page before rewriting title, meta and H1 around the top three terms. This week they watched it on a moving company's homepage that had six H1 tags, a 590-character meta description, and an H1 saying bay area when people search san francisco. One prompt fixed it.
@itsolelehmann [Claude Code]
https://x.com/itsolelehmann/status/2112301020009435363
A creator turned Claude Code into a free retention coach for Instagram, YouTube and TikTok. Screenshot the retention graph from the app, send Claude Code the video link and the image, and it downloads the audio, transcribes it with timestamps, lines the transcript up with the graph and quotes exactly what was being said at each big drop. It finishes with the one change that would keep the most viewers, and it installs whatever tools it needs along the way.
@migdoesit [Claude Code]
https://x.com/migdoesit/status/2112446732126638539
A developer gave their Claude-equipped Meta glasses a computer of their own using the Cua SDK. Saying hey Claude, use your computer and draft an email in Gmail sends the task to a separate machine rather than their Mac, and the glasses say in their ear when it is done. It runs on their local Claude Code with no API key, a small demo of agents getting their own desktops instead of borrowing yours.
@poolmoon6 [Claude Code]
https://x.com/poolmoon6/status/2112079299578679355
A Chinese user finally got Grok Bot to command the Claude Code, Codex, dsh and pi instances on their own computer, and stresses this is not just opening a Codex session in a terminal but two-way conversation, like Grok Bot talking to its own bots. They say there were many pitfalls and promise a full tutorial. It is one of several posts in the window treating a chat assistant as the dispatcher and local coding agents as the workers.
@fujitech_ai [Claude Code]
https://x.com/fujitech_ai/status/2112119948957147226
A Japanese engineer took a coding test for a side job, 30 minutes for one algorithm problem and one small ML exercise on public data, docs allowed but no AI, screen shared. The approach came quickly, but grasping unfamiliar data, looking up syntax and debugging did not fit in 10 to 15 minutes per problem, because for over a year nearly all their code has been written by Claude Code or Codex. They admit it stings, but argue that for experienced engineers who now delegate to AI, a better test is to allow AI and ask the candidate to explain the code's problems and design choices.
@kajikent [Claude Code]
https://x.com/kajikent/status/2112368799123132720
A Japanese manager wrote about why perfectionism and a Claude Code or Codex work style are a bad mix. There is always one more instruction you can give, and the uncertainty in agent output works on the brain like a small gamble, so without a self-imposed stop you become someone endlessly feeding context to AI. The old motto of always giving 120 percent only worked because human limits arrived early. Their fix is a deliberate 70 percent cap, which still gets two to three times the work done compared with before AI.
@bendee983 [Claude Code]
https://x.com/bendee983/status/2112535925926342805
A tech writer watched a product team expand with people who had near-zero experience and only used Claude Code, never looking at the repo, while the one software engineer was mocked as slow and an endangered species. Then the code grew into something even the agents could not handle, and the engineer was the one who got them unstuck. The story is anecdotal, but it is the concrete version of what several voices said this window: delivery still needs someone who can read the system.
@JFPuget [Claude Code]
https://x.com/JFPuget/status/2112118534129987833
A machine-learning researcher says AI agents now handle all of their training experiments, and wonders how much longer they will keep reading harness code. They are on Codex for now, put off by the Claude Code TUI rather than the models. The complaint that applies to every tool they tried: coding agents over-engineer, make changes nobody asked for that must be rolled back, and keep dead code for a compatibility that never mattered, which hurts readability and burns tokens. Their verdict is that software engineering is largely automated but not solved; you still have to drive.
@UPing123zzz [Claude Code]
https://x.com/UPing123zzz/status/2112574130235351190
A Chinese forward-deployed engineer says that with Codex and Claude Code a very impressive demo takes a few hours, but a demo and a delivery are different things. Take RAG: thousands of customer documents, updates, per-department permissions, wrong retrievals, made-up answers, unanswerable questions, and who is responsible when it breaks in production. The observation is that AI lowers the bar to build software while raising the bar for good engineers, and the gap is now understanding requirements, designing architecture, validating output and actually shipping.
@korenmiklos [Claude Code]
https://x.com/korenmiklos/status/2112639255872987156
An economist preparing a coding-agent demo for academic economists plans to tell the room to step outside the OpenAI and Anthropic bubble. Their credential is having tried Claude Code, Codex, Cursor, OpenCode, Hermes Agent, pi, Kilo Code, Factory, Kimi Code, T3 Code, Muse Code, Zed's Delta, Letta Code, omp, prime-agent and fx. They now work mostly in omp and Hermes, and their advice to anyone who has only seen Codex and Claude Code is to try OpenCode first.
@keitowebai [Claude Code]
https://x.com/keitowebai/status/2112516141465960767
A Japanese developer listed ten settings to keep information from leaking out of Claude Code: prefer auto over bypass mode, put pre-send actions on ask and anything off-limits on deny, launch from the project folder rather than home, scope API keys narrowly with short expiry and spend caps, deny reads of .env and credential folders, cut log retention from 30 to 7 days, review permissions monthly, put curl, wget and git push on ask for sensitive work, use only MCPs from clear publishers, and fence everything with the sandbox. The honest part is the footnote: they still launch from home and run full bypass on tasks that seem safe.
@theo [Claude Code]
https://x.com/theo/status/2112035326625730981
The founder of T3 Code shared usage data from their own client: for the last two weeks, more than 60 percent of T3 Code users who used Codex also used Claude Code. In other words, the majority of Codex users there are Claude Code users too. It is a small but real number behind the multi-agent habit the rest of this edition describes.
@smalkalbani [Claude Code]
https://x.com/smalkalbani/status/2112268475922358767
An Arabic-speaking creator was asked for a simple explanation of how an AI agent works and decided to answer with a video, giving Claude Code and Codex the exact same prompt. Claude took about 17 minutes and Codex about 8, and the author judged Claude's video the winner. The note that these explainer videos used to be exhausting to make is the real point.
@Ilya_Kuprov [OpenClaw]
https://x.com/Ilya_Kuprov/status/2112615765656453479
A researcher at the Weizmann Institute told OpenClaw's creator that it saves their life there and basically runs the lab by now. It is one line, but it comes from a scientific lab rather than a developer, at a moment when much of the feed is declaring OpenClaw over.
@surim0n [OpenClaw]
https://x.com/surim0n/status/2112456999782785030
A traveler tried to one-shot Korean arrival immigration forms for the whole family with several agents as the flight left the tarmac. Dot, Hermes and OpenClaw came close but refused sensitive steps like submitting, Muse could not find buttons, and Codex with 5.6 got there but asked for confirmation at every sensitive step. Only Instinct did it end to end, and the family went through Korean immigration without a hitch, though the author admits not being sure what it submitted was correct.
@ianlapham [OpenClaw]
https://x.com/ianlapham/status/2112383864563802203
A builder spent the past year making personal software factories and tried everything on launch day: OpenClaw, Hermes, every agentic IDE, Warp factories, Droid and every mix of cloud VMs and local tools. Nothing came close to a simple bot, Slack and a custom harness. Their recipe is to pipe in your context, develop good skills, build loops you like, verify on device and run nightly self-improvement, with Opus 5.5 as the sweet spot and Conductor for the hardest tasks. Everything else got abandoned.
@carlosazaustre [OpenClaw]
https://x.com/carlosazaustre/status/2112237349560000610
A Spanish developer had left their OpenClaw agent Tenacitas on autopilot for months because changes were hard to make. They migrated all of its skills, agents and memory to Hermes, trying it for the first time, and gave it GLM 5.3 and DeepSeek 4.1 as the brain through a subscription. The difference reminded them of when Opus could still be used inside OpenClaw. Tomorrow's daily tech news from the agent runs on the new brain.
@leploutos [OpenClaw]
https://x.com/leploutos/status/2112224926774047138
A French consultant who sets up agents for small businesses says the most common confusion is between model and agent. The agent is the harness, Hermes or OpenClaw, which keeps memory, reads mail and calls tools on a machine you control; the model is the brain it calls, by API or on your own GPUs. 99 percent of their clients run an agent on a VPS wired to models by API and none run a model locally. That separation is what makes switching easy: a better model comes out, you change one key, and the agent keeps its memory, tools and habits.
@michae1frank [OpenClaw]
https://x.com/michae1frank/status/2112200599999594665
A user moved their personal AI rig to 100 percent American open source, running Nemotron 3.5 Super VL locally on an Nvidia DGX Spark alongside OpenClaw and Codex. It took a while to get there. They call the resurgence of American open-source developers the most profound AI development of 2026.
@gabrielbuzziv [OpenClaw]
https://x.com/gabrielbuzziv/status/2112303725796864321
A Brazilian developer pulled an M1 Pro out of a box where it had sat for ten months, after putting off reformatting since the OpenClaw boom in March because setting up a machine from scratch used to eat an afternoon. This time Claude installed on the first try, got the list of programs needed and installed everything by itself. The machine now runs a self-hosted bot in place of Grok Bot.
@tarasshyn [OpenClaw]
https://x.com/tarasshyn/status/2112167451794219483
A SaaS marketer shared a visibility tactic: list your product on ClawHub, the official skill registry for OpenClaw agents, which carries domain rating 83. Write a SKILL.md describing what the product does, how to get an API key and which endpoints to call, then publish with the clawhub CLI; there is no review queue, only an automated security audit. Users install with one command, the page gives a dofollow link, and scrapers copy ClawHub skills onto directories you never submitted to.
🗣 User Voice
User Voice
The limit question flipped again. @marouane53 dropped a 200-dollar Codex plan when Opus 5.5 felt unlimited, and a few weeks later is hitting the cap on two Max subscriptions within three days. @KolinHuang says the same task has used about three times the quota since the weekend, and @ishuagra02 could run several Opus 5.5 Ultracode sessions on a 100-dollar plan at launch and now struggles to get past one, suspecting the desktop app update rather than a nerf. On the other side @SKatalystAI is running ten agents on each of three projects and still sits at 25 percent of the week. @atmoio asks what people do when the five-hour window runs out mid-feature, and answers that third-party clients are the only way through.
Switching tools still means starting over. @tongzhenbao is choosing between a second 100-dollar account and the 200-dollar tier, and notes that unlike Codex, switching Claude Code accounts loses the conversation and memory. @shiri_shh describes the loop of hitting the limit, moving to Codex, spending 20 minutes re-explaining the project, and moving back. That demand is why a memory layer that carries sessions across 20-plus harnesses launched with a seed round the same weekend.
The quality complaint is now about excess, not errors. @JFPuget says every agent tried over-engineers, makes unrequested changes that need rolling back and keeps dead code for compatibility nobody needed. @ilyamiskov, who polishes graphics pixel by pixel, spends twenty times longer fixing a photorealistic bookshelf than drawing it by hand in Framer. @UPing123zzz puts it as a demo in hours, a delivery never, without someone who owns permissions, updates and failure modes.
The cost is also personal. @sasaki445 spent a month raising Claude Code like a colleague and now feels stress severe enough to worry about health, and wants to stop using AI. @kajikent caps their own effort at 70 percent because one more instruction is always available and agent output feels like a small gamble.
Permissions friction cuts both ways. @gregpr07 is being asked to approve actions while running bypass permissions, @allgarbled finds auto mode over-restrictive with no working override, and @lucasmeijer reports that subscription-powered prompts are rejected unless they start with the Claude Code system line, despite docs saying the SDK and third-party apps can use the subscription.
The limit question flipped again. @marouane53 dropped a 200-dollar Codex plan when Opus 5.5 felt unlimited, and a few weeks later is hitting the cap on two Max subscriptions within three days. @KolinHuang says the same task has used about three times the quota since the weekend, and @ishuagra02 could run several Opus 5.5 Ultracode sessions on a 100-dollar plan at launch and now struggles to get past one, suspecting the desktop app update rather than a nerf. On the other side @SKatalystAI is running ten agents on each of three projects and still sits at 25 percent of the week. @atmoio asks what people do when the five-hour window runs out mid-feature, and answers that third-party clients are the only way through.
Switching tools still means starting over. @tongzhenbao is choosing between a second 100-dollar account and the 200-dollar tier, and notes that unlike Codex, switching Claude Code accounts loses the conversation and memory. @shiri_shh describes the loop of hitting the limit, moving to Codex, spending 20 minutes re-explaining the project, and moving back. That demand is why a memory layer that carries sessions across 20-plus harnesses launched with a seed round the same weekend.
The quality complaint is now about excess, not errors. @JFPuget says every agent tried over-engineers, makes unrequested changes that need rolling back and keeps dead code for compatibility nobody needed. @ilyamiskov, who polishes graphics pixel by pixel, spends twenty times longer fixing a photorealistic bookshelf than drawing it by hand in Framer. @UPing123zzz puts it as a demo in hours, a delivery never, without someone who owns permissions, updates and failure modes.
The cost is also personal. @sasaki445 spent a month raising Claude Code like a colleague and now feels stress severe enough to worry about health, and wants to stop using AI. @kajikent caps their own effort at 70 percent because one more instruction is always available and agent output feels like a small gamble.
Permissions friction cuts both ways. @gregpr07 is being asked to approve actions while running bypass permissions, @allgarbled finds auto mode over-restrictive with no working override, and @lucasmeijer reports that subscription-powered prompts are rejected unless they start with the Claude Code system line, despite docs saying the SDK and third-party apps can use the subscription.
📡 Eco Products Radar
Eco Products Radar
Grok Bot: the assistant most often named as a dispatcher that hands work to local Claude Code and Codex sessions.
Codex: the default second agent, as reviewer, implementer, or the place people go when Claude limits run out.
Beacon: open-source memory layer that turns coding sessions into skills across 20-plus harnesses, launched with a 5.5 million dollar seed and reposted all weekend.
Sunra MCP: image, video and audio models inside Claude Code and other assistants at the makers' API prices, promoted by many accounts at once.
DeepSeek V4.1 Flash: the cheap model under the harness, in harness benchmarks and plan-in-Claude, write-in-DeepSeek setups.
Hermes: the agent people migrate OpenClaw setups to, and half of the usual Hermes or OpenClaw question.
Jev: small decision model pitched in many model-routing templates; mostly promotion rather than first-hand use.
Blender: driven through Python or MCP for ads, short clips and game assets with no video model.
Unreal Engine plugin: Epic's official Claude Code plugin that controls the live editor over MCP, with Unity shipping its own the same weekend.
graphify: turns a folder into a knowledge graph the agent can query instead of reopening files.
MiniMax M3.1 Flash: newly available in OpenClaw and on Ollama, and used inside the Claude Code harness for cheap 3D scenes.
Devin Fusion: Opus 5.5 planning plus GPT-6 Luna executing, reported above Claude Code on a coding-agent index at lower cost per task.
durable-harness: Vercel's open-source way to make Pi, Claude Code, Codex or OpenCode resume after failures.
t3code: third-party client praised for multi-machine links and cross-agent threads.
Cua: SDK for giving agents their own computers, used from Claude Code and even smart glasses.
Obsidian: the vault behind most second-brain and project-memory setups.
Grok Bot: the assistant most often named as a dispatcher that hands work to local Claude Code and Codex sessions.
Codex: the default second agent, as reviewer, implementer, or the place people go when Claude limits run out.
Beacon: open-source memory layer that turns coding sessions into skills across 20-plus harnesses, launched with a 5.5 million dollar seed and reposted all weekend.
Sunra MCP: image, video and audio models inside Claude Code and other assistants at the makers' API prices, promoted by many accounts at once.
DeepSeek V4.1 Flash: the cheap model under the harness, in harness benchmarks and plan-in-Claude, write-in-DeepSeek setups.
Hermes: the agent people migrate OpenClaw setups to, and half of the usual Hermes or OpenClaw question.
Jev: small decision model pitched in many model-routing templates; mostly promotion rather than first-hand use.
Blender: driven through Python or MCP for ads, short clips and game assets with no video model.
Unreal Engine plugin: Epic's official Claude Code plugin that controls the live editor over MCP, with Unity shipping its own the same weekend.
graphify: turns a folder into a knowledge graph the agent can query instead of reopening files.
MiniMax M3.1 Flash: newly available in OpenClaw and on Ollama, and used inside the Claude Code harness for cheap 3D scenes.
Devin Fusion: Opus 5.5 planning plus GPT-6 Luna executing, reported above Claude Code on a coding-agent index at lower cost per task.
durable-harness: Vercel's open-source way to make Pi, Claude Code, Codex or OpenCode resume after failures.
t3code: third-party client praised for multi-machine links and cross-agent threads.
Cua: SDK for giving agents their own computers, used from Claude Code and even smart glasses.
Obsidian: the vault behind most second-brain and project-memory setups.
Comments