Super User Daily: 2026-11-21
Thursday, November 19 was IPO day for Anthropic, and the most useful posts were about a quieter question: how do you know the agent actually did the thing. A Japanese game-media editorial team went back through 153 failure notes Claude Code wrote for itself and found 82% of them came from runs where every exit code said success. A robotics workshop measured Claude Code driving a two-armed robot and found it was thinking more than half the time and moving barely a third of it. One builder got a whole operating system to boot on a 2012 ThinkPad only because the OS was designed from day one so the agent could read its own screen through a serial port. The other strand is people who are not writing software at all: a $10 UGC ad for a Texas burrito chain, a WeChat sticker set for a cat, a vocal cover with Synthesizer V, talking illustrations animated with code, and an MMORPG opening cinematic scripted shot by shot. The pattern on both sides is the same. The work goes to the agent, and the human job turns into building the check.
@fax_v2_0 [Claude Code]
https://x.com/fax_v2_0/status/2123366040872558602
A builder wrote an entire operating system with Fable 5.2 through Claude Code and got it booting on a real 2012 ThinkPad T430. It is not a Unix: there are no files, a copy-on-write B-tree database sits at the core, and you type SQL at the prompt, so even the task list is a table you query. About 25,000 lines of Rust with zero external crates, its own TCP/IP stack, SATA and Intel Ethernet drivers, 229 unit tests and 326 whole-system checks in QEMU. The rule that made it work was set on day one: the OS must be drivable through a serial port in a machine-readable format, so the agent could boot it, type commands, read the screen and check its own work. Real hardware still got the last word, because the network card received everything and sent nothing for five trials until one register was written bit by bit instead of whole.
@ai_chuhai [Claude Code]
https://x.com/ai_chuhai/status/2123350283338006824
A Japanese game-media editorial team has had Claude Code building Godot and Unreal projects since June: four Godot projects, 7,878 commits, 81% co-signed by Claude. Every time Claude hit a time-wasting trap it wrote itself a failure note, and the team went back through all 153. In 125 of them, 82%, the exit code and the test results showed no failure at the time, and the process never stopped. The worst case was 40 Godot stats scripts running in parallel, all returning 0 and marked as passed, while every log said SCRIPT ERROR because the code called a String constructor that does not exist in Godot 4, so every stat stayed at zero. The fix was to judge completion by several signals at once: ERROR lines in the logs, whether the expected files exist, pixel reads from screenshots on the production renderer, and per-match logs from automated games, with feel and fun left to a human.
@junfanzhu98 [Claude Code]
https://x.com/junfanzhu98/status/2123234896730792301
A co-organizer of the CoRL 2026 agentic robotics workshop posted the most measurement-heavy talk of the day: Claude Code on Opus 5.5 and Codex on GPT-6 Astra driving a wheeled two-armed robot through MCP tools, with no object detector, on fruit into a basket, a bowl onto a rack and opening a door. The robot was actually moving only 20 to 36% of the time, and thinking took 53% of task time for Opus. After learning a skill, Opus finished a task in about 2.4 to 5.1 minutes for $1.2 to $2.0, while Astra took about 10 minutes for $4 to $7, and both hit 100% on fruit and door but 0% on the bowl. Giving the agents more tools did not help, and rehearsing in a simulation first lifted the bowl task only to 25%. The safety lesson is blunt: code has git revert and robots have no undo, one agent asked to retract 48 seconds after it sensed high force, and stopping has to take milliseconds without waiting for the model.
@leploutos [Claude Code]
https://x.com/leploutos/status/2123349060073521540
A solo founder shipped a dictation app for kids in six weeks: a native iPhone app, a native Android app, a server and five builds sent to Apple and Google, without writing the code personally. Several Claude Code sessions run in parallel, and the founder only coordinates. Every repo has a CLAUDE.md with product rules (a mistake in a generated dictation is unacceptable, no data about the child leaves the phone), the build and test commands to run before saying done, and the traps already hit, like purchase tests that always fail on the first run of a fresh simulator. Next to it sits a decisions file with more than a hundred dated choices and their reasons, so a new agent knows why cross-platform was rejected without being told. Everything goes through a GitHub issue with a platform label and a version milestone, which is what made the parallel sessions stop stepping on each other.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2123204689982296185
A marketer made a 27-second AI UGC ad for a Texas burrito chain for about $10 in credits, with Claude Code running the whole thing from one terminal. It pulled the chain's live Meta ads (one holiday gift card promo, no food in it) and scraped their last 48 Instagram posts, 77 images, so the generated food is their actual steak burrito in the red tortilla rather than a stock burrito. It wrote a 66-word, 9-shot script, placed an avatar in a booth in their real store with the surf mural from Instagram, cloned a voice, and rendered talking clips and five food shots including the queso pour. Then whisperx checked every spoken word against the script, zero wrong, and ffmpeg cut it on the word timings with captions and an end card. Restaurant UGC is the most formulaic ad format there is, which is exactly why it suits an agent.
@DomainMaster [Claude Code]
https://x.com/DomainMaster/status/2123465841672904839
A developer fixed 18 bugs in a browser robot football game with one real prompt and a mixed team of agents. Codex coordinated, Grok Bot and Muse playtested the live site and documented bugs with repro steps and recordings, and Claude Code investigated the causes and wrote the fixes. Codex merged duplicate reports, routed evidence to Claude, followed up when tests stalled and arranged retests, for example when Grok Bot said the match clock was still hard to read after the first fix. The human pressed Tab and Enter once to approve merging and deploying three PRs. Of the 18 issues, 11 were reported only by Grok Bot, 2 only by Muse, 4 by both and 1 was found by Claude during investigation, which is a decent argument for using more than one playtester.
@nikarpit [Claude Code]
https://x.com/nikarpit/status/2123301636638752939
A product designer turned founder says the team barely opens Figma anymore, because the whole design workflow moved into Claude Code over a few months. Designers now spend most of their time building prototypes, managing the design system and working directly on the frontend, so the job title effectively changed to design engineer. Figma is left as a scratchpad for dumping ideas, the design system lives in a designsystem.md, and skills tell agents what to check and review before anything ships. Subagents explore several design directions in parallel so the team can see whole flows side by side. The point is not speed alone: you can build a version, feel it and ship it without explaining every detail you imagined to someone else.
@sidvalecha_ [Claude Code]
https://x.com/sidvalecha_/status/2123243160134681086
As San Francisco flooded, one builder made a live flood map in about two hours with Claude Code and Grok. Grok vision reads street signs and storefronts in flood photos posted on X to work out where each one was taken. Those points are then laid over National Weather Service alerts, tide data and SFPUC and FEMA risk zones. It is a small project, but it shows the useful shape of a two-hour civic tool: a vision model to geolocate messy public photos, official layers for context, and an agent to glue it together while the event is still happening.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2123320618661929399
A Chinese builder ran 12 animation skills against the same job, six animations for one bakery, and compared them with plain Claude Code. Without any skill, Claude Code got five of six right for $1.77 in total. Most skills got about the same number right but were slower and more expensive, only two were faster than no skill (both for web motion), and every single one cost more. The three worth installing were narrow ones: sprite-gen for pixel sprite sheets, img2threejs for 3D characters and SeeCode for flow animations. The full ranking covers 79 animation skills, and the headline lesson is that a skill has to beat the bare model on your own task, not on its README.
@1red2black [Claude Code]
https://x.com/1red2black/status/2123337784748294326
A Russian-speaking developer shared a recipe against over-engineering: a separate session renamed bigboss that every other agent must report to. Claude Code sessions on the same machine can message each other, so the architect and programmer sessions are told to go to bigboss before changing the plan, changing the list of product features, or asking for work to be accepted, and to keep reworking until bigboss is satisfied. The boss runs on a model with good taste but low reasoning effort, deliberately not too clever and allergic to complexity. The prompt casts it as a blunt, always-right founder figure, so the agents stop gold-plating because someone with veto power keeps asking whether the feature is needed at all.
@zasuko_michiksa [Claude Code]
https://x.com/zasuko_michiksa/status/2123295105453068654
A Japanese creator made a talking explainer animation of themselves and a mascot cat entirely in code with Claude Code, with no video generation model. The illustration is split into eyes, mouth, hair and other parts as a line-free vector puppet, body motion comes from NVIDIA's Kimodo, which turns text into skeleton movement, and lip sync follows the loudness of the voice automatically. Blinking, breathing and eye movement are automatic too, and the cuts, diagrams and subtitles are all code. Because every part is a parameter rather than a generated frame, a fix means changing one value, not rerolling the whole clip.
@ivy432hz [Claude Code]
https://x.com/ivy432hz/status/2123398117072764957
A hobbyist bought Synthesizer V 2 and a voicebank and did a self-cover of a song originally made with Suno, with every step done by Claude Code. The human side was just repeated requests along the lines of change this part. Claude Code handled the vocal synthesis setup and even took noisy, very personal taste notes without complaint, and the creator also staged the cover with an official MMD model. The slightly off notes stayed in on purpose because they sounded cute. It is a good example of Claude Code as a studio assistant for someone who knows what they want to hear but not how to drive the tools.
@atom63_ [Claude Code]
https://x.com/atom63_/status/2123327722101637432
A creator made a WeChat sticker set and a promo video for their cat with one setup. Claude Code on a Mac mini orchestrates the work, Codex generates the images, the MiniMax API animates them and code assembles the video, while the creator directs everything from a phone. There is no agent framework at all. The standard operating procedure grew as the work went on, kept as plain docs in a repo, which is probably the most copyable part: the process is just files the agent reads next time.
@kokushing [Claude Code]
https://x.com/kokushing/status/2123345150956978267
A Japanese indie developer made three opening cinematics for their own MMORPG with Tripo Studio and Claude Code, about 3 minutes 20 seconds, 20 seconds and 22 seconds long, in a post sponsored by Tripo. Tripo models the goddess, the cave and the village, Codex drew the goddess design, Blender handles assembly and bones, and Claude Code writes the script, camera, staging, capture and edit, all running live in Godot rather than as video files. The script is a timetable of camera, character, light, screen and music, cameras are placed relative to anchor points like the cave exit, and stills at each timestamp are lined up and reviewed before checking in game. Fixes came from that review, like pulling the camera back when the waking shot cut off the head. The whole method is written up as a guide Claude Code reads before each job.
@MengTo [Claude Code]
https://x.com/MengTo/status/2123355714693521766
A designer built a turn-based three.js game with Fable 5.2 in two days and tried something new for the characters. The designer gave Claude Code one picture of each character plus the Hi3D MCP and just watched. Claude Code uploaded each image, checked the cost first, started all four jobs at once, kept polling until they finished and downloaded each model as a GLB. The human part was the pictures, and everything between a 2D sketch and a usable 3D asset was handled by the agent through one MCP.
@BuildFastWithAI [Claude Code]
https://x.com/BuildFastWithAI/status/2123256021531701383
A builder pitted Fable 5.2 in Claude Code against GPT-6.2 Astra in Codex on the same job: model a Porsche GT3 RS in Blender. Claude Code was set to keep iterating until a critic approved the result or got it close enough to the reference, and it burned through an entire week's quota doing so. Codex took a completely different approach, three hours and just two messages, though it ignored the instruction to remove the plates. Fable won clearly on detail and Astra on speed. The useful lesson is the critic loop itself: an open-ended keep going until approved instruction is exactly how a weekly limit disappears in one task.
@turingou [Claude Code]
https://x.com/turingou/status/2123276231554548169
A Tokyo-based founder ran out of Claude Code quota again this week while traveling, after doing over 90% of the work from a phone against Claude Code's cloud environment. The case for it: inference and compute sit in the same place, so the cloud environment handles 80 to 90% of automation work without permission cards, and results do not have to cross the ocean to a laptop and back. The founder used to run builds and development on a company server in Tokyo, but driving that server from the cloud now asks for more permission prompts than working directly in the cloud, so most automated development moved into Claude Code's built-in cloud projects. The init script restores logins and pulls backups whenever the shared machine is rebuilt, and read-only keys live in the environment's secret store. The longer-term plan is to turn the Mac minis and Mac Studio into CI machines and treat the personal computer as a screen.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2123246130771173563
One strawberry took two weeks, 3,958 instructions and 17,453 tool calls from Claude Code agents. The result is a procedural Blender plugin with about 7,700 geometry nodes that generates the fruit. It is an extreme example of what agents can grind through when the output is a node graph rather than a mesh: thousands of small, checkable edits that no human would want to place by hand. It is also a useful cost reference for anyone planning procedural asset work, because the effort sits in iteration, not in a single clever prompt.
@elionaz [Claude Code]
https://x.com/elionaz/status/2123275572453294150
A developer built agenterm, a macOS app that runs Claude Code, Codex, Gemini and any other agent CLI side by side, grouped by project. Each agent gets its own git worktree, so parallel sessions never edit the same checkout. Sessions outlive the app, so closing the window does not kill a long run, and the app pings you when an agent needs input. It is free and MIT licensed, and it targets the most common complaint of the week: agents silently waiting for approval while you are looking at something else.
@bojie_li [Claude Code]
https://x.com/bojie_li/status/2123335123232031209
A Chinese AI founder shared a small trick for long tasks in Claude Code and Codex: make the agent do the work in a subagent and keep the main session for talking with you. By default the work happens in the main session, so a quick question often waits minutes while it wraps up the current step, and the transcript fills with command details you do not care about, which makes scrolling back to a conversation from hours ago painful. With chat in the main session and execution in subagents, the interface gets much cleaner. As the author puts it, congratulations, you just turned Claude Code into the main-session-chats, background-executes architecture that personal agents use.
@binghott [Claude Code]
https://x.com/binghott/status/2123531340725506552
An ads operator had written off Meta's MCP with Claude as hot garbage the first time around. This week the same setup with Claude Code on Opus 5.5 uploaded videos and duplicated existing ads to launch new ones, easily and accurately, with no third-party tools in between. That is the kind of repetitive Ads Manager work agencies bill for every month. The shift worth noting is not the model alone but the official connector getting good enough that the middleware layer disappears.
@ayumi_t820 [Claude Code]
https://x.com/ayumi_t820/status/2123269429198999810
A Japanese creator's YouTube video titled video editing with only Claude passed 1,000 views, the first of their horizontal videos to do so. The method uses Claude Code on Opus 5.5 with Remotion, a simple standing character illustration and a prepared template to produce explainer videos. The template carries the visual identity and the agent fills in the content, which keeps each new video cheap. The small milestone matters because the audience showed up for the method itself.
@adrienrusso [Claude Code]
https://x.com/adrienrusso/status/2123421657909883151
A French creator asked followers how to make an animated video and ended up producing it with Claude Code, Remotion, ElevenLabs, OpenAI image models for flare and sunburst effects, and royalty-free music. The result came out very well in what was almost a single prompt. The stack is the one that keeps showing up for motion video: Remotion for the timeline, ElevenLabs for the voice, an image model for the effects, and Claude Code writing the glue code.
@alex_borto [Claude Code]
https://x.com/alex_borto/status/2123360887168000055
A French developer built a website using only AI, with Claude Design for the look and Claude Code for the build. It took time on purpose, because the goal was not to ship the same generic page everyone has seen, and the PageSpeed score came out excellent. The catch was search: the AI completely ignored SEO and GEO fundamentals until it was pushed hard to add them. Content updates currently go through the cPanel API, and the open question is how management holds up once the site grows beyond a few pages.
@nyxienyxie [Claude Code]
https://x.com/nyxienyxie/status/2123460854519664985
Mastra open-sourced the factory that maintains its own framework, and last week its agents merged 238 PRs next to 630 from humans. The post's point is that the PR count is the wrong thing to copy. WorkOS built the standard setup first, sandbox plus model plus repo with prompt in and PR out, and it was no different from engineers running Claude Code on their laptops. What changed results was putting the team's process inside the factory: explicit gates for intake, triage, plan, build, review and done, review comments wired back into the running session, and a markdown wiki with source citations as company memory. Mastra even closed the repo to outside PRs, and contributors now send detailed repros instead, because when agents write the code the scarce input is a good problem statement.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2123498797636735174
Zapier grades every hire on four levels of AI skill, and using ChatGPT to draft product specs only gets a PM to level two. A product writer showed three levels of personal work to Zapier's CEO on a podcast: basic ChatGPT use was rated Capable, a one-person PM OS built in Claude Code was rated Adoptive, and a full Company OS was rated Transformative, because transformative means you are the first of your kind doing the thing. The top cells in the rubric are telling, since both end up used by the whole company: eval-first product development and a feature on a fine-tuned model that opens a new pricing tier. So the writer built a Company OS in 11 days and rolled it out to the team. The interview is now show and tell, and the bar is work that lands in other people's hands, not an afternoon saved.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2123441314192818600
A growth marketer released an open-source AI video editor that a coding agent drives, a hybrid of Descript, Frame.io and After Effects. It does word-by-word transcription of an hour of footage in about two minutes, speaker diarization, audio-based sync of every camera angle, screen recording and mic, and cloud GPU rendering of an hour-long video in minutes. All the editing happens through Claude Code, Grok Bot or Codex, which read the transcripts, pick takes, cut ums and dead air, and switch to whoever is talking. The part that makes it a team tool is multiplayer review: comment on a frame or range, start a comment with @agent and it becomes a task the agent claims, edits, re-renders as v2 and replies to with what changed and the timecodes.
@911_takuya [Claude Code]
https://x.com/911_takuya/status/2123565080629637617
A Japanese developer added MCP endpoints to several self-built systems so they can all be operated from Claude Code. Claude Code now acts as the hub that manages those systems together instead of each one needing its own dashboard. The author calls it ridiculously convenient and recommends it. It is the in-house version of the plugin trend: you do not wait for vendors to ship connectors when the systems are your own, you give each one a small MCP surface and let one agent sit on top.
@FinansowyUmysl [Claude Code]
https://x.com/FinansowyUmysl/status/2123407695596191916
A Polish user left a work computer on the desk, plugged in and with sleep disabled, installed Tailscale on it and on an iPhone, then installed T3 Code. Now the user can work remotely from the phone, anywhere, on their own computer and with their own Claude Code setup. The reply thread is full of alternatives, from tmux on a VPS to a Telegram bot that launches Claude Code on a laptop and Claude Code's own remote option. The common need is clear: keep the agent running on hardware you control and reach it from your pocket.
@cxjwin [Claude Code]
https://x.com/cxjwin/status/2123256763340444007
A Chinese user ran a checkup report through Mirobody, an open-source health records project, entirely on a Mac with no API key: llama.cpp for a 2B model, glm-ocr for reading images, on an M4 Pro with 48GB. The first upload of a 14-page PDF imported 33 rows and none from the three lab pages, because those pages were scanned images with a text header, and the rule skips OCR on any page with a text layer. Re-uploaded as images, all 58 lab values came in and matched the original, with three small slips like an extra digit in one reference range, and 20 of 58 rows were mapped to standard LOINC codes. One detail stood out: the label for creatinine was ambiguous in LOINC, so it left the code empty instead of guessing. For the Claude Code step the author deliberately switched to the project's demo account, since data queried over MCP goes to Claude's servers, and with only one record it said plainly that no trend could be seen.
🗣 User Voice
User Voice
People want to see what the agent saw, not just whether it stopped. @lenny_enderle found Claude Code reading files itself was fooled by a forged record 1 time in 180, while a cheaper Haiku subagent reading the same files was fooled 75 times, and asked whether anyone knows which model their subagents run on. @stas_sorokin_ ran a booby-trapped setup script past both agents: Claude Code never ran it, but the session simply ended with a safeguard notice and never explained what was wrong, which leaves a user likely to run the script by hand. @checkpoint_gg_ reminded everyone that viral paste-this-into-your-agent prompts install whatever the script says on the day you run it, and @ClaudeCodeLog noted that 2.1.329 dropped the built-in rule to pause before high-blast-radius actions.
Effort is a budget dial, and users have started reading it that way. @Arindam_1729 pointed to Artificial Analysis numbers where Fable 5.2 in Claude Code scored 58 on high at $7.51 a task but 56 on max at $44.22 and 2.2 hours, @ScarletKc found Claude Code in the most expensive tier of a third-party harness benchmark, and @iamleannmuller cut a chat from 194,400 tokens to 19,800 with /clear and still ran out two days before reset, because subagents, noisy test runs and unused skills drain the limit.
Context does not survive the jump between tools. @DanielSmidstrup switches to Codex when Claude Code hits its limit and then explains the whole project again from zero, and @buildwithquinn burns 170k tokens at every session start across three repos, to which @mattpocockuk answered that the codebase is your memory and the setup needs pruning, not another skill.
Mixing harnesses is popular and legally murky. @tamo3dess found Antigravity needs hand-written hooks for permission splits that Claude Code handles with an allow and deny list, @RetroOptions dropped a plan to add Google's top plan after learning that calling Antigravity from Claude Code is not allowed, and @gaojingyu tried Google's model plus Claude Code's harness plus GPT for tests and concluded the human had become the harness.
The backlash against setup hype is growing. @leandronsp gets plenty done with tmux and one or two Claude Code sessions and calls fleet orchestration pure hype, @Emil__Lio argues the most hardcore builders mostly just dictate into Claude Code or Codex, @tokengobbler saw 78% of a London meetup pick Claude Code, and @StatsLime still wants an interface that makes it easier to review what the agent wrote.
People want to see what the agent saw, not just whether it stopped. @lenny_enderle found Claude Code reading files itself was fooled by a forged record 1 time in 180, while a cheaper Haiku subagent reading the same files was fooled 75 times, and asked whether anyone knows which model their subagents run on. @stas_sorokin_ ran a booby-trapped setup script past both agents: Claude Code never ran it, but the session simply ended with a safeguard notice and never explained what was wrong, which leaves a user likely to run the script by hand. @checkpoint_gg_ reminded everyone that viral paste-this-into-your-agent prompts install whatever the script says on the day you run it, and @ClaudeCodeLog noted that 2.1.329 dropped the built-in rule to pause before high-blast-radius actions.
Effort is a budget dial, and users have started reading it that way. @Arindam_1729 pointed to Artificial Analysis numbers where Fable 5.2 in Claude Code scored 58 on high at $7.51 a task but 56 on max at $44.22 and 2.2 hours, @ScarletKc found Claude Code in the most expensive tier of a third-party harness benchmark, and @iamleannmuller cut a chat from 194,400 tokens to 19,800 with /clear and still ran out two days before reset, because subagents, noisy test runs and unused skills drain the limit.
Context does not survive the jump between tools. @DanielSmidstrup switches to Codex when Claude Code hits its limit and then explains the whole project again from zero, and @buildwithquinn burns 170k tokens at every session start across three repos, to which @mattpocockuk answered that the codebase is your memory and the setup needs pruning, not another skill.
Mixing harnesses is popular and legally murky. @tamo3dess found Antigravity needs hand-written hooks for permission splits that Claude Code handles with an allow and deny list, @RetroOptions dropped a plan to add Google's top plan after learning that calling Antigravity from Claude Code is not allowed, and @gaojingyu tried Google's model plus Claude Code's harness plus GPT for tests and concluded the human had become the harness.
The backlash against setup hype is growing. @leandronsp gets plenty done with tmux and one or two Claude Code sessions and calls fleet orchestration pure hype, @Emil__Lio argues the most hardcore builders mostly just dictate into Claude Code or Codex, @tokengobbler saw 78% of a London meetup pick Claude Code, and @StatsLime still wants an interface that makes it easier to review what the agent wrote.
📡 Eco Products Radar
Eco Products Radar
Codex: the default second seat, as reviewer, coordinator or head-to-head rival on the same task.
Gemini 4 Neon and Antigravity: launched the day before, the main comparison target, with a recurring debate about whether its subscription may be driven from other harnesses.
Grok Bot: playtester, orchestrator and the personal agent people keep looking for cheaper replacements of.
Muse: second playtester and light personal agent next to Claude Code.
Remotion and FFmpeg: the code-to-video layer under almost every agent video workflow.
ElevenLabs and Whisper: voice and transcript checks in automated video production.
Blender and Blender MCP: 3D modeling for cars, machines, procedural assets and game scenes.
Godot: the game engine behind cinematics and the 153-failure-note study.
T3 Code: the app people use to run several harnesses and reach them from a phone.
Tailscale: the private network that makes a home machine reachable for remote Claude Code.
Fable 5.2 and Opus 5.5: increasingly chosen by effort level and cost per task, not by default.
Volcano: Kong's agent cloud, promoted heavily on launch day as a Claude Code plugin.
Linger: an open-source local personal agent that uses Claude Code or Codex as its engine.
Codex: the default second seat, as reviewer, coordinator or head-to-head rival on the same task.
Gemini 4 Neon and Antigravity: launched the day before, the main comparison target, with a recurring debate about whether its subscription may be driven from other harnesses.
Grok Bot: playtester, orchestrator and the personal agent people keep looking for cheaper replacements of.
Muse: second playtester and light personal agent next to Claude Code.
Remotion and FFmpeg: the code-to-video layer under almost every agent video workflow.
ElevenLabs and Whisper: voice and transcript checks in automated video production.
Blender and Blender MCP: 3D modeling for cars, machines, procedural assets and game scenes.
Godot: the game engine behind cinematics and the 153-failure-note study.
T3 Code: the app people use to run several harnesses and reach them from a phone.
Tailscale: the private network that makes a home machine reachable for remote Claude Code.
Fable 5.2 and Opus 5.5: increasingly chosen by effort level and cost per task, not by default.
Volcano: Kong's agent cloud, promoted heavily on launch day as a Claude Code plugin.
Linger: an open-source local personal agent that uses Claude Code or Codex as its engine.
Comments