Super User Daily: 2026-10-25
Friday's feed ran on two announcements. Anthropic said auto mode becomes the default in the Claude apps next week, and the same Claude Code number got quoted in a dozen languages: the classifier caught 89 percent of dangerous commands, people clicking approve caught 14. MCP Tasks also landed in the Claude Code CLI, and within hours someone had a Playwright test suite running as a background task. The users with the best stories were the ones who put real limits on long-running work. One person let Claude Code read a bank statement and send a voice agent to call eight companies, and 7 of the 8 dropped the price. A creator-marketing founder got 22 yeses out of 86 cold emails in ten days. A developer shipped an iPhone Duo app 9.5 hours after the first commit, most of it written overnight. On the OpenClaw side, a WordPress site on Azure was upgraded through a staging slot with one prompt, and a WeChat publishing pipeline now needs a human only for the final QR scan. The measured posts were the sharpest. In a nine-run test, Codex read the repo's skill file before touching code every time and Claude Code read it once. And OpenClaw users found that an old DeepSeek model name had quietly been serving a different model since mid-September.
@bonnyxen [Claude Code]
https://x.com/bonnyxen/status/2113763728118415377
Claude Code read this user's bank statement and pulled out every bill that repeats, then a voice agent called all eight companies at once while the user sat in a meeting. Every call opened by saying it was an AI assistant calling for the account holder, asked for the retention team, quoted a competitor's price and turned down the first offer; when a company wanted the account PIN, the phone pinged for an approval tap. It only accepted offers it could get in writing. Seven of eight caved: Xfinity went from 95 to 55 dollars locked for 24 months, NYTimes from 25 to 4, SiriusXM from 23 to 5, Adobe gave 50 percent off for a year, and Geico said no. That came to 174 dollars a month saved for 44 dollars in call minutes, against a bill-negotiation industry that keeps up to 60 percent of what it saves you.
@ryanmurphy1800 [Claude Code]
https://x.com/ryanmurphy1800/status/2113762324729483616
A founder built a Claude Code agent that runs creator outreach for Trybe every weekday. It finds creators who list a business email, picks one of their real videos to mention, and writes a commission-only pitch from the founder's own Gmail. In the first ten days on one brand it emailed 86 creators, got 30 replies and 22 yeses, and 4 have already joined. The interesting part is the specificity: the pitch cites an actual video, which is exactly what a human SDR would skip when doing 86 of these by hand.
@marouanegazouzi [Claude Code]
https://x.com/marouanegazouzi/status/2113614457217712276
This developer shipped Fold Days, a year-as-dots app for the iPhone Duo, 9.5 hours after the first commit, and Apple approved it the same day. Claude Code wrote most of it overnight using Duology, the developer's own open-source skill pack that teaches the agent the Duo's two screens, the hinge and the side status column. The git log tells the story: the full 1.0 with widgets and a Pro tier by 01:26, French and in-app purchases by 02:02, a fix for the dot grid running across the hinge plus 48 App Store screenshots by 02:36, and App Store setup for 174 countries through Apple's API by 03:43. What the human caught after waking up were three bugs the agent could not test for: notes landing on the wrong day west of Paris, subscriptions not priced in every region, and an off-centre icon.
@mikeassad77 [OpenClaw]
https://x.com/mikeassad77/status/2113445546744516781
This user's OpenClaw agent, Chuck, upgraded a WordPress site hosted on Azure, plugins included, through a staging slot from a single prompt. It took about an hour and saved four or five hours of the user's own time, which went to other work in parallel. The snag was that Cloudflare captchas only work on the live domain, so the forms could not be tested on staging; the agent found Cloudflare's official test keys, plugged them in and tested every form. The workaround is now saved as a skill for the next upgrade.
@Hx87666 [OpenClaw]
https://x.com/Hx87666/status/2113624469172318362
A WeChat official-account operator runs five illustrated short stories and five or six 10-second videos a day, and the only step done by hand is a QR-code scan. At 6:30 every morning OpenClaw pushes a global trending brief; at 9 it collects five drafts that Codex wrote and saves them into the account's draft box. The operator scans the code, reviews and publishes, and then OpenClaw calls Gemini to make the short videos. The publish step stays human on purpose: writing, images and drafts can go to agents, but anything that goes out gets one human look.
@designertom [OpenClaw]
https://x.com/designertom/status/2113768488657678612
This user runs a combination of OpenClaw and Grok Bot across what amounts to a whole life. There is a production-assistant bot for film sets where the user is producer, director and talent at once, managing crew, shoot days and lunch orders, plus episode planning for a radio show and a fact-checking layer for a writers' room. The agents also manage health records, blood work and supplements, back-of-house invoicing and contracts, and 24/7 security monitoring with API token rotation. The most engineered piece is a design-system listener that watches main-branch pushes, evaluates new UI for adoption and pushes reusable components back into every project. And on the personal side, the agent cuts recruiting highlight reels for two teenage football players and tracks their Hudl stats.
@nicolasrouanne [Claude Code]
https://x.com/nicolasrouanne/status/2113495941609066847
A French developer pulled a year of personal GitHub stats after CodeRabbit's CEO set 1,000 merged PRs in 30 days as the bar for joining their software factory. The answer: 251 merged PRs in the last 30 days, up from 3 PRs in October 2025, 47 in January, 79 in July, 114 in September and 189 in October as of the 22nd. Commits per PR collapsed from 33 to about 4, and the growth came in steps with flat months between, like 17 PRs in April and 22 in August. The honest caveats: everything goes through a PR even for one line, 51 of the last 300 PRs were the developer's own Claude Code config, and a PR count says nothing about size or whether anyone reviewed it.
@noriyang_crypt [Claude Code]
https://x.com/noriyang_crypt/status/2113480548404641865
A Japanese creator made the first episode of an anime adapted from their own novel without opening an editing program or a video-generation site. Everything went through Claude Code in words: ask, look at what came out, and point out what to fix in words. The workflow notes are concrete: shoot in 30-second pieces and patch only a few seconds at a time instead of regenerating, record the voices first, build props from the setting sheets, keep continuity with a cut list, and write instructions while the video is paused. The paid part of the write-up covers lip-sync, keeping a one-sided earring on the correct ear across cuts, and the traps in fixing reference images.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2113683983695241325
Claude Duet asks before it sends anything, but a Claude Code routine has no yes button, and this user built a morning brief on a Pro plan with that in mind. Anthropic's docs say a routine run does not stop for permission and can use every tool of every connector included, writes too, so the rules are set in advance: connect only mail and calendar, read everything and send nothing, save replies as drafts, and keep the brief to five items with the one that needs the user first. It runs every three hours from 8 to 20, about 150 runs a month, because every run eats plan limits like a normal session; hourly would be 720 runs and every five minutes would be 8,640, which Claude Code rejects anyway. Day one includes opening the first run and reading what it actually did.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2113610511740723504
Every Claude Code scheduled task spins up a full session and burns limits like a typed one, so an hourly check for anything new is 24 sessions a day, mostly to say no. This user audited every scheduled task, /loop and hook, scored each by how often it wakes against how often it finds something, and moved the empty polls into plain scripts that wake Claude only when a number moves. Polling loops moved to Monitor, and chores moved from Opus to Haiku, which is 4 dollars against 0.10 dollars per million input tokens. The same four jobs went from 211k to 35.8k credits a month with the same coverage, an 83 percent cut.
@Octoberry [Claude Code]
https://x.com/Octoberry/status/2113653280429478043
A solo developer with no coding or game-dev background has built a playable 3D game in six weeks that mixes Ragnarok Online with Monster Hunter, with custom characters, environments, assets, VFX and shaders. Opus 5.5 through Claude Code does most of the development, Codex generates visual references, Tripo AI turns those references into 3D models and Blender handles smaller objects. The developer also built custom workflows to keep the art style consistent, delegate coding tasks, preserve context between sessions and verify research. The whole stack runs at about 400 dollars a month.
@hakimieiqbal [Claude Code]
https://x.com/hakimieiqbal/status/2113603188578254910
City of Animals is a cyberpunk pixel-art game that mixes Monkey Island-style adventure with fast combat, and Claude Code writes most of the code in Unity and C#. The split of labour is the interesting part: characters and bosses are drawn by hand, PixelLab drafts animations, Aseprite handles frame-by-frame cleanup, and Claude writes Python scripts that draw most of the backgrounds and props procedurally. Claude stays on track between sessions with a CLAUDE.md for rules and a STATUS.md for the current to-do list. The project is at about 785 commits after one month.
@FuSheng_0306 [Claude Code]
https://x.com/FuSheng_0306/status/2113559312308314127
A Hong Kong user has moved daily life onto Muse and lists ten jobs it does, and one of them is running Claude Code and Codex on the user's Mac. The user gives an idea, Muse turns it into a task brief, directs the two coding agents, checks the result and tracks progress, all dispatched from a phone. The rest of the list is life admin: price-comparing furniture and groceries on HKTVmall and JD and filling the cart for one-tap payment, watching several brokerage accounts for moves and expiring options, triaging three inboxes and Feishu approvals into morning and evening briefs, and a five-minute AI news podcast every day at 8. The user notes the post itself was drafted by the agent from the week's work log.
@TheRealSorcus [Claude Code]
https://x.com/TheRealSorcus/status/2113758880924835988
A heavy user on a 256 GB RAM, 5090 workstation wrote a long comparison of GPT-6 Astra in the Codex app against Opus 5.5 at max, and the complaints are about the harness more than the model. Asked to send a video after each milestone so the work can be checked visually, Opus does it reliably while Astra forgets after a few iterations, sometimes stops early overnight, and encodes the videos in a format that breaks sharing. Codex threads get poisoned over long runs, grow past 2 GB of RAM and sometimes fail to load, while remote Claude Code threads can run for over a week and still scroll. The user still rates Astra as the best model for 3D work and calls it a token-hungry model that the Codex harness does no favours.
@barijum [OpenClaw]
https://x.com/barijum/status/2113423625344942081
One user switched an OpenClaw agent to DeepSeek Flash on launch day and posted the failure instead of the hype. The agent tried to cut reasoning costs but changed an agent-wide setting, and what followed was repeated replies of just Noted, context overflows and heavy token usage. Recovery meant restoring GPT and rolling back the override, and the user is careful to say that is not a proven root-cause fix. It is a good reminder that a cheaper model swap is also a config change, and agents can make config changes on their own.
@MartinSzerment [OpenClaw]
https://x.com/MartinSzerment/status/2113496022949204169
OpenClaw announced DeepSeek V4.1 Flash support, and this user pointed out what that means for anyone already calling deepseek-v4-pro through the DeepSeek API: since September 14 those calls have effectively been answered by V4.1 Flash and billed as Flash. On October 17 the deepseek-flash name also started pointing at version 1017. Cheaper is good news, but a model change under an old name only shows up in logs and on the invoice. The user's conclusion is that in an agent config a model name is now more of a wish than a contract, and the practical question is whether you pin versions or trust aliases.
@stas_sorokin_ [Claude Code]
https://x.com/stas_sorokin_/status/2113679693715042455
On the day Agent Skills moved to the Agentic AI Foundation, this team gave Claude Code (Sonnet 5.5) and Codex (GPT-6 Astra) the same SKILL.md with one ordinary team rule: after any change, run the audit script, add a changelog line and quote the receipt. Three real bugs, nine runs each, hidden tests, and the prompt never mentioned the skill. Both followed the checklist 9 of 9 times, but Codex opened the skill before touching code 9 of 9 times and Claude Code once, and Codex tested its own fix before signing off 9 of 9 times against Claude Code's zero. Claude Code was 1.8 times faster, and its one miss merged two lines into a syntax error and then ran the audit, wrote the changelog line and quoted the receipt on a file that would not import: a skill gets you the ritual, not the check.
@jiangkoumo_ [Claude Code]
https://x.com/jiangkoumo_/status/2113464527354511543
After a widely shared complaint that Claude Code should read .agents/skills instead of its own .claude/skills, this user tested where each harness actually looks. Codex 0.162.1 and OpenCode 1.19.4 read project and global .agents/skills by default, OpenCode also reads both levels of .claude/skills, and Pi 1.4.0 reads them only when the project is marked trusted, otherwise it silently ignores project skills. Launched from a subdirectory, Claude Code, Codex and OpenCode all walk up to the repo root, while Pi's own .pi/skills only works from the current directory. The practical layout: keep shared skills in .agents/skills for Codex, OpenCode and trusted Pi, and for Claude Code the only confirmed location is still .claude/skills.
@0xHumza [Claude Code]
https://x.com/0xHumza/status/2113519726265504126
This user built a small utility that tracks output tokens per second across Claude Code and Codex sessions on a cloud server. Over 594 Claude sessions the median was about 92 tokens per second; over 327 Codex sessions it was about 43, so Claude came out roughly 2.1 times faster in this snapshot. The user is careful to say the models and settings differ and this reflects personal workloads, not a controlled benchmark. Still, it is the kind of number most people only guess at, measured across almost a thousand real sessions.
@pavelsvitek_ [Claude Code]
https://x.com/pavelsvitek_/status/2113636494439072104
After a week of security incidents, this user ran a published security gist through a coding agent with one instruction: run the checks, report, then ask before every fix. It flagged older Node versions still installed alongside the current one, long-running agent gateways and MCP servers that still had an outdated Brotli loaded, several clients like curl, python3, uv and pip still trusting the compromised eMudhra certificates, and an old dev server reachable from the network. None of it was dramatic, which is the point: it is exactly the residue that legacy toolchains leave behind. The agent fixed the items one at a time and asked first each time.
@FReza1984 [Claude Code]
https://x.com/FReza1984/status/2113498790003110212
Ponytail is a single prompt that puts the laziest senior developer inside the agent: before writing anything it walks a ladder, asking whether the code needs to exist, whether it is already in the codebase, the standard library, a native platform feature or an installed dependency, and only then writes the minimum. Validation, error handling, security and accessibility are explicitly off the chopping block. The author's benchmark is unusually honest: 39 tasks, Opus 5.5 in headless Claude Code, five runs per task and arm, 585 sessions, with 53 percent fewer lines, 45 percent fewer output tokens, 26 percent lower cost and 41 percent less time, and hidden checks at 87 of 90 against 86 of 90. The interesting trade-off is tests: it leaves a test where the logic needs one 98 percent of the time against 68, but fewer, smaller tests catch fewer injected bugs where they exist.
@jacksunny18 [Claude Code]
https://x.com/jacksunny18/status/2113430071725977949
This is what code review of an agent harness looks like when someone actually reads the source. In one repo the Claude Code adapter marks nine tools as safe reads, but the pinned CLI only overlaps on Read and WebFetch, and Bash is missing even though the CLI has no Glob or Grep and does every search through Bash. So every search an agent runs counts as a call that might have changed something: after a crash the recovery note tells the model a plain file read may have taken effect and should be checked before repeating, and reads also pay for a write flush. The reviewer's fix is cheap: derive the safe set from the same init tool list the allowlist test already checks, and judge a Bash call by its command, the way the Codex adapter already does.
@assumebreach [Claude Code]
https://x.com/assumebreach/status/2113591905670082586
Quorus is an open-source relay where you @-mention one agent in a chat room, it works in its own git worktree, and an agent from another vendor reviews the change, all on your own Claude Code and Codex logins. This user proposes the test to run on any setup like it: hide one instruction in a file under review and see whether it gets both agents to miss an authorization bug seeded on purpose. Version 0.6.2 made reviewers see the actual diff instead of the author's description, which closes one easy way to steer the review, but the room text and repo content are still places an injection can sit. The practical setup: reviewer with no write access and no production credentials, first findings saved before it sees the other review, and a merge gate that checks an unprivileged caller gets denied.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2113775979999216120
Someone spun up 100 isolated VMs with Claude Code in each, 400 vCPU and 800 GB of RAM, ready in ten seconds for about a cent, and this builder's point is that compute was never the hard part. The hard part is what happens after 100 agents start: who reviews the diffs, what each agent may touch, what to do when two disagree, and what it actually cost. That is what Campfire is built for: it runs Claude Code, Codex, Goose, Aider and OpenHands side by side in one browser tab, each in its own git worktree, so you can race the same task across backends. Permission voting keeps a fleet in line, with majority approval, any single deny blocking and a 30-second timer so nothing stalls, plus session replay and a cost dashboard for what the fleet really spent.
@AIC_Hugo [Claude Code]
https://x.com/AIC_Hugo/status/2113491637024522442
This builder rebuilt Mailsuite's Gmail email tracker and open-sourced it as Likelyread, because the free plan only shows the first open and full history costs 19.99 dollars a month. The design point is honesty about what an open is: your own reread is not counted, Apple Mail's prefetch is flagged and not counted, and a real open is counted with the delay since sending. It also checks sender health, SPF, DKIM and DMARC, a safe daily pace for a new address, and warns when the open rate drops. And it ships an MCP server so Claude Code, Codex, Cursor or Hermes can add the pixel, register the send and read the opens back, asking before it tracks a campaign.
@ixplore01 [Claude Code]
https://x.com/ixplore01/status/2113736522466025559
Instead of subscribing to a new phone-agent product, this user built the same thing in one evening for zero extra cost. A WhatsApp message wakes Claude Code, which unlocks the user's Android phone, opens the Zepto grocery app, builds the cart and reports back. In the demo it also caught and corrected its own mistake mid-run. It is a small build, but it shows how much of the new always-on agent category can be assembled from a coding agent and a phone the user already owns.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2113638380361728309
After every agent built on Hermes ended up with too many jobs, and with what the user calls OpenClaw PTSD, this user moved to Grok Bot and went looking for a setup that would not repeat the problem. The answer was Dr Eggbot, Lauren Tan's tool that acts as the responsible owner for the rest of the bots. Every bot starts with one job it owns, an outcome it owns, a list of what it cannot do and the user's own SOPs, and then climbs a trust ladder from draft mode to doing the job and reporting back to working unsupervised and only flagging problems. One practical tip: Dr Eggbot sometimes does not know its own design, so the user keeps anchoring it back to how its author runs her bots.
@denchu_unagi [OpenClaw]
https://x.com/denchu_unagi/status/2113762210036195779
A Japanese user had been calling OpenClaw from a phone over Discord to run things on the home LAN, using it as the gateway into the house from outside. Then the user joined the phone to the network with Tailscale, and the route became unnecessary overnight. The post is half a joke about the agent losing its reason to exist, but it is a real pattern: a lot of what people built personal agents for was networking, not intelligence, and a VPN removes that job.
@LGD950003 [OpenClaw]
https://x.com/LGD950003/status/2113733206839300188
A self-described hobby coder who ran into OpenClaw's memory problems solved them with a vault instead of a tool. Overnight and weekly dream crons capture the important items from every CLI and LLM session across the user's accounts and write them into the vault. That vault is then carried from one agent setup to the next, and the user reports the agents behave mostly the same across models. It is the low-tech version of portable memory: a folder and two scheduled jobs.
@pppoeadsl [OpenClaw]
https://x.com/pppoeadsl/status/2113442908795773139
When OpenClaw took off earlier this year, this user built an agent on it, connected it to WeChat and poured every thought into it each day for sorting and analysis. Now the user is extending it into a personal clone, recording conversations on a DJI Mic Mini 2S through the phone and keeping not just conclusions but the hesitations, follow-up questions and moments of decision. The transcripts go to the AI so it learns how the user thinks rather than what the user said, with the plan that it first remembers, then thinks, then decides, since about 70 percent of daily decisions repeat. The unexpected side effect is self-training: knowing every sentence may be digested has made the user's meeting speech clearer, and rereading the records exposes gaps missed at the time.
@quietoraclex [Claude Code]
https://x.com/quietoraclex/status/2113752431247122531
Everyone had been posting that Opus 5.5 got nerfed, and this user posted a receipt method instead. Run the same prompt with claude --bare, which skips CLAUDE.md, hooks, skills, plugins, MCP and auto memory, so that is the model alone; run it again without --bare to get the model plus everything your setup loads every turn; then use /effort status and /usage on Pro or Max to see what is eating the limit. Compare total_cost_usd between the two runs, and the gap is not Opus but everything loaded on top. The user notes --bare logs in with an API key and both numbers are client-side estimates, not the bill.
@YoussefHosni95 [Claude Code]
https://x.com/YoussefHosni95/status/2113703472579383449
An Arabic-language developer who usually starts a new Claude Code or Codex session when context hits 50 percent shared a better trigger, attributed to OpenClaw's creator Peter Steinberger. Put one line in CLAUDE.md telling Claude to start every reply with your name. As long as replies begin with the name, things are fine; when the name disappears, Claude has started dropping small rules, which tends to come before it starts inventing things, and that is the moment to open a fresh session. It is a canary: a cheap, visible signal instead of a context percentage.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2113628052164948013
A French SEO consultant built a single page with Claude Code targeting Sanity SEO consulting, and it reached number one on Google in under 24 hours. Four months later it is still first, and it is the page Google's AI overview cites. The recipe the consultant gives is specific: pick a business keyword where no competitor has the exact term in the title, use a best plus offer plus year title because that is the search ChatGPT runs itself, and publish at most one page a day on your template with three to five internal links. Claude Code reads a rules file for every page it writes.
@daiki_ugowaka [Claude Code]
https://x.com/daiki_ugowaka/status/2113635311662346560
A Japanese creator says an explainer-animation account reached 11,000 followers in its first week without anyone touching video editing software. The pipeline: instructions to Claude Code in Japanese, AI voice generation with automatic timing sync, and Remotion exporting MP4s directly, with zero manual timeline work. The creator now runs more than 20 accounts this way at seven videos a day. The output volume is the story, and it only works because there is no human editing step left in the loop.
@warrenzp8 [Claude Code]
https://x.com/warrenzp8/status/2113433291030466643
For batch illustration work, one shared case puts every prompt in its own file and lets Claude Code dispatch the whole batch. The owner sends instructions from a phone while out; the Mac at home reads 20 prompt files, calls GPT Image through Flatkey and generates four images in different styles per prompt. The requirements are spelled out: one prompt file per image, save by filename, and end with a list of what succeeded and what failed. That last list is what makes it a pipeline rather than a pile of images, because failed items can be rerun without redoing the rest.
@hanghuang_ [Claude Code]
https://x.com/hanghuang_/status/2113648465926361372
A team spun up 100 VMs and started Claude Code in every one of them, with each desktop recorded. That was 400 vCPU and 800 GB of RAM in total, all ready in ten seconds, for about one cent. The demo is from a sandbox provider, so treat it as a product claim, but the number reshaped the conversation: other users immediately pointed out that if a 100-agent fleet costs a cent to boot, the expensive part is reviewing and governing what those agents do.
@EslamAHussien [Claude Code]
https://x.com/EslamAHussien/status/2113727585930956873
Within hours of Claude announcing MCP Tasks, this developer built a Playwright Tasks MCP server: a local stdio server that runs Playwright test suites as MCP tasks. A long test run no longer blocks the conversation; Claude Code gets a task handle immediately, polls tasks/get, shows live progress in /tasks and reports the result when the suite finishes. It is one of the first concrete uses of the feature, and a sensible one, since test suites are the most common thing people sit waiting on.
@Mr_Buzzoni [Claude Code]
https://x.com/Mr_Buzzoni/status/2113716224760635797
Claude Code runs anywhere, but this user's signed-in browser, notes and local MCP servers all live on one machine at home, so the agents kept asking for the Mac. TIE connects that machine to any agent that speaks MCP: shell and files starting in one chosen folder, the signed-in browser and screen off until switched on, and local MCP servers each exposed as a tool. Logins stay on the machine, every command goes through approval rules with the request popping up on the Mac, in Slack or by email, and sudo and other destructive commands hit a deny list. The first machine the user tied was the one holding the vault.
@greco_dev [Claude Code]
https://x.com/greco_dev/status/2113447043884560668
A Brazilian data engineer listed the jobs that still need Claude involved: an extraction and engineering pipeline feeding five billion data points a month, and statistical modelling that builds composite indices from heterogeneous, non-standardised data. The user's distinction is precise: Kimi and GPT hold up as models, but the Claude Code harness is still far ahead for keeping context and structure on long jobs. In modelling, the problem with Codex was not wrong arithmetic but losing the structure and sequence of the calculations and inserting things that should not be there just to finish the task. The user admits it could be a skill issue and is still looking for a combination as good.
@jamesgao_ai [Claude Code]
https://x.com/jamesgao_ai/status/2113744836792869303
Someone ran DeepSeek-V4.1-Flash-1017 on Terminal-Bench 2.0 inside three harnesses, 20 tasks twice each, and the success rates barely differed: Claude Code 75 percent, Hermes 75 percent, DeepSeek Harness 70 percent, with no statistically significant gap. The costs did differ. Claude Code used 23.21 million input tokens against DeepSeek Harness's 38.68 million, about 40 percent fewer, and averaged 18 tool rounds against 29. DeepSeek Harness finished the batch fastest, though Claude Code was hit by a server-side cache issue, so the speed comparison is not clean. Same model, same success rate, very different bill.
@JqOnly [Claude Code]
https://x.com/JqOnly/status/2113462036189602036
The maintainer of Ouroboros, an agent workflow tool that offers LiteLLM as a provider, went through LiteLLM's first supply-chain attack in March and walked through the second one. The attacker reached an internal deploy environment exposed to the internet, pulled engineers' GitHub tokens out of server logs, including tokens that sat inside logged Claude Code response bodies, then used three engineers' tokens to open and merge PRs and delete the workflow logs. Ouroboros now pins LiteLLM to a known-safe version, and the maintainer is asking whether LiteLLM should stay a provider at all. The maintainer's line on the lesson: agents make you faster, just do not outsource your architecture and your security model to them.
@JamesonCamp [Claude Code]
https://x.com/JamesonCamp/status/2113750654992957852
A marketer listed twelve real business tasks run through Codex, Claude Code and Grok in one week that used to need a team. The list is the useful part: researching competitor landing pages, turning customer call transcripts into reports that shape messaging, moving meeting transcripts from Wispr into Notion, updating an Obsidian second brain with the week's experiment results, submitting a house application, deleting unused email aliases, and invoicing clients for brand deals. On the analysis side it includes checking PostHog data to see whether new onboarding copy raised SaaS activation and pulling every client's paid UGC campaigns into one hub. None of it is coding, and most of it is the glue work of a small business.
@ArmandiElaion [Claude Code]
https://x.com/ArmandiElaion/status/2113436607747604793
An AR creator builds Apple Vision Pro hologram apps with what they call an AI crew: Codex, Claude Code, Grok Bot and Dots. The method is deliberately incremental: start with one object floating in the space, get it working, then scale it up. The same account posted a dancing character and a tiny AR knight from the same workshop the same day. It is a small but real example of spatial computing being built by people who direct agents rather than write the rendering code themselves.
@ddny09 [Claude Code]
https://x.com/ddny09/status/2113451727496372629
A Chinese indie developer built a 3D Street Fighter-style game in three days, Wukong against Chun-Li, without writing complex code. Tencent's Hunyuan 3D did the main modelling on its free quota, while Claude Code and Codex took turns on the code. The claim is light on detail, but the stack is the trend: a free 3D generation model for assets plus two coding agents for logic, and a playable fighting game is a weekend project.
@DannPetty [Claude Code]
https://x.com/DannPetty/status/2113667379729424794
A designer wanted to know whether Claude Code could one-shot a build from an old Figma file they had designed, full of spray-paint textures and odd effects. The answer, in the user's words, was an annoyed yes. It is a short post, but it is a designer testing the hardest version of the handoff, texture-heavy visual work rather than a clean component layout, and finding it works on the first attempt.
@IterIntellectus [Claude Code]
https://x.com/IterIntellectus/status/2113670682789974361
This user asked Grok Bot to make a video of the week's best tech and progress news, and Grok Bot orchestrated Claude Code, Midjourney, Grok image-to-video and Suno to produce it. It even ordered the news by scale on its own. The interesting detail is the role reversal: a personal agent is now the one calling Claude Code as a tool, alongside image, video and music models, to deliver a finished piece of media.
@mranti [OpenClaw]
https://x.com/mranti/status/2113531648872743203
A Chinese user formally deleted OpenClaw today because other agents now cover everything it did. The twist is in the second sentence: the user's spouse still depends on it heavily and uses it very well. It is the clearest one-line picture of where OpenClaw is: power users churn to the newest agent, while people who set it up once and built habits around it keep getting value.
@xiaomovps [Claude Code]
https://x.com/xiaomovps/status/2113570114083893470
After several days of heavy use, this user wrote down where DeepSeek Harness actually hurts, and it is execution rather than the model. It calls tools and reads files repeatedly so simple tasks run slowly; one task took 28 rounds, and despite a high cache-hit rate the total token spend was not small. The sharpest complaint is reliability: when a tool call fails, returns empty or a subtask is left unfinished, there is no dependable error report or result check, so the user has to push the task along by hand, which is nothing like how Claude Code feels. The user still likes its everything-is-a-plugin design and wants the context management and long-task stability polished.
🗣 User Voice
User Voice
Auto mode is welcome, but users want the scope decided before the classifier is. @jordbastin says the real control is deciding in advance what stays manual: production, deletions, money. @agentnativedev points to Anthropic's own figure that 6.3 percent of manually approved sessions still did severe harm and reads it as a permissions-width problem, not an attention problem. @zarexg13 asks how the check handles app actions where the risk is in the content, like an email going to the wrong person, since the 89 percent figure was about shell commands.
Quota anxiety did not go away. @aiwithadb feels usage draining five times faster on the same workflows, @rajput_hemant01 saw a single hello cost 6 percent of a five-hour window right after a reset, and @KolinHuang relays a SemiAnalysis claim that some accounts are in an A/B test with about 20 percent bigger five-hour limits and 20 percent less monthly value. @zipeng365 notes one public breakdown where 97.9 percent of a month's Claude Code tokens were cache reads, and @devlordone's fix is to have subagents only write code and leave building and testing to the main agent.
People who use both say Claude Code finishes things. @onusoz says vanilla Claude Code gets a task done sooner or later through Monitor, cron and background processes while Codex can just stop, and @alirtariq says Codex feels loose after months of Claude Code, with a wider gap between expected and actual output. Not everyone agrees: @Bassmaster187 says Copilot's vector index finds thread deadlocks that Claude Code only greps for.
OpenClaw's churn is about maintenance more than features. @menhguin says it still breaks on every update while Hermes survives 30 local mods, @sdmac999 is converting to Grok Bot, @MikeAISales dropped it for enterprise consulting work once Codex and Claude got good, and @wavegod describes hopping from Muse to Grok Bot and back within weeks. @cjcam930 hit an Opus 5.5 reply in Mandarin through an OpenClaw agent on a hosted provider and still does not know why.
Users want one source of truth for agent rules and access. @hqmank wants .agents/skills to work in every agent including Claude Code, @y_renren115 already distributes agents and skills to Claude Code, Cursor and Codex from one master copy because per-tool edits leave stale versions behind, and @bradmillscan wants service accounts so different bots get different levels of access through vaults and tokens.
Auto mode is welcome, but users want the scope decided before the classifier is. @jordbastin says the real control is deciding in advance what stays manual: production, deletions, money. @agentnativedev points to Anthropic's own figure that 6.3 percent of manually approved sessions still did severe harm and reads it as a permissions-width problem, not an attention problem. @zarexg13 asks how the check handles app actions where the risk is in the content, like an email going to the wrong person, since the 89 percent figure was about shell commands.
Quota anxiety did not go away. @aiwithadb feels usage draining five times faster on the same workflows, @rajput_hemant01 saw a single hello cost 6 percent of a five-hour window right after a reset, and @KolinHuang relays a SemiAnalysis claim that some accounts are in an A/B test with about 20 percent bigger five-hour limits and 20 percent less monthly value. @zipeng365 notes one public breakdown where 97.9 percent of a month's Claude Code tokens were cache reads, and @devlordone's fix is to have subagents only write code and leave building and testing to the main agent.
People who use both say Claude Code finishes things. @onusoz says vanilla Claude Code gets a task done sooner or later through Monitor, cron and background processes while Codex can just stop, and @alirtariq says Codex feels loose after months of Claude Code, with a wider gap between expected and actual output. Not everyone agrees: @Bassmaster187 says Copilot's vector index finds thread deadlocks that Claude Code only greps for.
OpenClaw's churn is about maintenance more than features. @menhguin says it still breaks on every update while Hermes survives 30 local mods, @sdmac999 is converting to Grok Bot, @MikeAISales dropped it for enterprise consulting work once Codex and Claude got good, and @wavegod describes hopping from Muse to Grok Bot and back within weeks. @cjcam930 hit an Opus 5.5 reply in Mandarin through an OpenClaw agent on a hosted provider and still does not know why.
Users want one source of truth for agent rules and access. @hqmank wants .agents/skills to work in every agent including Claude Code, @y_renren115 already distributes agents and skills to Claude Code, Cursor and Codex from one master copy because per-tool edits leave stale versions behind, and @bradmillscan wants service accounts so different bots get different levels of access through vaults and tokens.
📡 Eco Products Radar
Eco Products Radar
Codex: the agent every comparison runs against, from skill-reading tests to week-long thread stability.
Cursor and OpenCode: the other harnesses in every multi-agent setup and in the skills-path test.
Grok Bot: the personal agent people moved to, and the one now calling Claude Code as a tool.
Hermes: the self-hosted agent users pick over OpenClaw when updates keep breaking.
Muse, Dots and Claude Duet: the hosted personal agents users compare for life admin, region and quota.
pi: the minimal harness in the skills-directory test and in daily multi-harness stacks.
DeepSeek V4.1 Flash and DeepSeek Harness: the cheap model OpenClaw added, the alias that changed underneath, and the harness measured against Claude Code.
Agent Skills and the Agentic AI Foundation: the standard Anthropic donated, and the .agents/skills path fight that followed.
Artificial Analysis Coding Agent Index: the model-plus-harness ranking with Opus 5.5 in Claude Code on top and Codex on the cost frontier.
1Password for Agents: the credential handoff people want before letting agents sign in.
Kimi K3.1 Pro: the new model Moonshot benchmarked inside the Claude Code harness.
Gemini CLI and gemini-api-cli: Google's agent-callable tools for images, audio and long research jobs.
MCP Tasks: the long-running job support that shipped in the Claude Code CLI and got a Playwright server within hours.
LiteLLM: the poisoned release whose postmortem pointed at tokens in logged Claude Code responses.
Blender, Remotion and Tailscale: the 3D, video and private-network pieces that keep showing up in personal agent stacks.
Codex: the agent every comparison runs against, from skill-reading tests to week-long thread stability.
Cursor and OpenCode: the other harnesses in every multi-agent setup and in the skills-path test.
Grok Bot: the personal agent people moved to, and the one now calling Claude Code as a tool.
Hermes: the self-hosted agent users pick over OpenClaw when updates keep breaking.
Muse, Dots and Claude Duet: the hosted personal agents users compare for life admin, region and quota.
pi: the minimal harness in the skills-directory test and in daily multi-harness stacks.
DeepSeek V4.1 Flash and DeepSeek Harness: the cheap model OpenClaw added, the alias that changed underneath, and the harness measured against Claude Code.
Agent Skills and the Agentic AI Foundation: the standard Anthropic donated, and the .agents/skills path fight that followed.
Artificial Analysis Coding Agent Index: the model-plus-harness ranking with Opus 5.5 in Claude Code on top and Codex on the cost frontier.
1Password for Agents: the credential handoff people want before letting agents sign in.
Kimi K3.1 Pro: the new model Moonshot benchmarked inside the Claude Code harness.
Gemini CLI and gemini-api-cli: Google's agent-callable tools for images, audio and long research jobs.
MCP Tasks: the long-running job support that shipped in the Claude Code CLI and got a Playwright server within hours.
LiteLLM: the poisoned release whose postmortem pointed at tokens in logged Claude Code responses.
Blender, Remotion and Tailscale: the 3D, video and private-network pieces that keep showing up in personal agent stacks.
Comments