Super User Daily: September 1, 2026
The story this week wasn't a new feature, it was a billing announcement that read like a magic trick: a "permanent 25% increase" to Claude Code weekly limits that, once the temporary 50% summer boost expires on September 14, actually lands users about 17% below where they are today. The timeline did the math and got loud. But underneath the quota drama, the more interesting signal is what people are actually doing with these agents. Non-coders are shipping internal business systems in a couple of weekends. Marketers run entire content operations off one terminal. And a growing crowd has stopped arguing about which model is smartest and started asking a different question: which harness wraps it best. Here's what real users built.
@santtiagom_ [Claude Code]
https://x.com/santtiagom_/status/2094075173632618863
A friend who runs a small merchandising business had spent years managing quotes, invoices, and sales in a pile of Google Docs. He downloaded Claude Code, described the business he knew cold, and two to three weeks later had a working internal system: quote products from the site, log invoices, register sales, everything in one place. He can't program. He just knew the problem well enough to explain it and correct it as it went. The one wall he hit was the plumbing around code itself, GitHub, versioning, domains, deploy, which is still confusing for someone who never wrote software.
@iamAlexTurnbull [Claude Code]
https://x.com/iamAlexTurnbull/status/2094088753283305787
A pointed counter to the vibe-coding hype from someone twelve months into building an AI-native Zendesk with ten engineers. His whole team codes heavily with AI, but he insists the code was always the easy part. The hard part is understanding the product deeply enough to know what to build: 50+ core requirements calls with support teams, and features that aren't independent. Change custom fields and you touch permissions, automations, routing, reporting, AI context, imports. You can vibe-code a throwaway WordPress app, he says, but you cannot sit down and vibe-code the next Zendesk.
@marshssg22 [Claude Code]
https://x.com/marshssg22/status/2094194224900043156
Built a full pest-control ad campaign on a brand-new ad account in under 30 minutes, start to finish, entirely in Claude Code. That included a lead form with qualifying questions, a campaign with two ad sets targeting exact zip codes with CBO, eight image ads, 26 ads total, and daily Slack reporting. His estimate for doing the same thing by hand: two to four hours. This is the kind of non-coding output that keeps showing up, an agent used as a media buyer, not a code generator.
@charliejhills [Claude Code]
https://x.com/charliejhills/status/2094054950619979831
A marketer, not a developer, who says Claude Code has run almost his entire content business for 180 days. He lays out the eleven things that cover 80% of it: plan mode, mid-session replan when output degrades, one CLAUDE.md per platform, automemory that stops repeating a mistake after one correction, a goal finish-condition, custom skills that turn terminal jargon into a browser page, and a handover.md written at 50% context so he can clear the session without losing any learning. The takeaway he keeps hammering: this isn't just for engineers.
@Feroliver19 [Claude Code]
https://x.com/Feroliver19/status/2094001077829972185
A non-technical founder deleted his Framer and Bolt accounts and rebuilt the company's entire website in Claude Code in a week, without pulling the engineering team off their work. He frames it honestly: it feels normal in startup circles, but for 99% of businesses this is still wild, one non-technical person shipping a production site solo.
@KevTheHermit [Claude Code]
https://x.com/KevTheHermit/status/2094160406935638494
A security researcher pointed Claude Opus and a custom research harness at the freshly announced PaperCut NG exploit to see if it could reverse-engineer the patch, and how fast. The run: 10 prompts, 293 tool calls, 224 million tokens, roughly 90 minutes, one guardrail. Claude took the patch, identified three critical vulnerabilities, and built a full-chain proof of concept resulting in unauthenticated code execution, then found a bypass for the released patches. This is authorized security work, but it's a stark look at how fast the offensive-research loop now closes.
@cryptojezuz [Claude Code]
https://x.com/cryptojezuz/status/2093979074972823637
Used to spend 4 to 6 hours every release doing manual regression checks across API endpoints. Now he drops a testing block into CLAUDE.md and runs a slash command before merging: run the regression suite against all existing endpoints, flag breaking changes to request/response shape even when tests pass, generate a compatibility report. Last week it caught a "minor" validation update returning 400s where it used to return 200 with a warning, a break that would have hit old clients in production. His real point isn't the speed, it's that with the friction gone he actually runs it every single time.
@mfpiccolo [Claude Code]
https://x.com/mfpiccolo/status/2094003705418789237
A clean articulation of why tabs full of agents isn't multi-agent orchestration. In his team's internal harness, Pi, Claude Code, Codex, and even VS Code are workers on one engine. A trigger wakes Pi, hands it the workspace, lets it do its part, then an event triggers Claude Code or Codex to continue. The interesting bit isn't running every agent CLI, it's that the harness can orchestrate them, with agents triggering, calling, and handing work to each other rather than being a nicer tab multiplexer.
@Michaelzsguo [Claude Code]
https://x.com/Michaelzsguo/status/2094045584327856557
Finally closed a 24/7 agent loop. He swapped the supervisor over his Omarchy Pi agent to Claude Code, which uses its own background-wait feature to check on the Pi agent's progress every 15 to 20 minutes. When it hits a problem, a cc-reviewer agent pings him for approval directly through the Claude phone app, so he got the request while reading a book. The setup, Herdr plus Claude Code plus the Claude app, means the Pi agent keeps working, Claude Code keeps supervising, and a human gets pulled in only when needed.
@alexhillman [Claude Code]
https://x.com/alexhillman/status/2093907627394601320
Replaced Claude Code's auto-compact with his own hook-based handoff workflow. Instead of compacting the same session, his meta-harness catches when the context window hits a threshold, generates a handoff to his spec, spins up a clean session, reads the handoff, and continues, and it starts the handoff before it's needed so it's fast. The handoff itself is generated by a cheap mix of code and Haiku. Driving from his meta-harness instead of Claude Code directly, he says sessions feel infinite.
@sora_biz [Claude Code]
https://x.com/sora_biz/status/2094032417522892986
Claude Code and Codex can each message across their own sessions, but not across each other, so he built a bridge harness. Left pane Claude Code, right pane Codex. He handed over a spec and said "implement it, get Codex's approval, ship to preview," and the whole loop, implement, request review, Codex checks against the spec and approves, publish preview, ran with zero manual steps. His deeper motivation is visibility: as everything goes cloud-agent, he wants to see with his own eyes that the agents are actually talking, not just claiming to.
@nifinet [Claude Code]
https://x.com/nifinet/status/2094154220698214557
Made a 68-second product launch film with one line of prompt, 38 minutes to first cut, zero After Effects. The workflow is the real lesson: first give Claude Code a skill that tears any launch video into an ingredients list without imitating it, have it study the best films and write teardowns to disk, build a throwaway practice film to leave behind a working Remotion project and render scripts, and put your brand palette on disk (Claude lifted his from the live site's computed styles). Every effect has to be achievable in CSS, and each scene is its own component so you can redo scene nine without touching the rest.
@TylerG_Capital [Claude Code]
https://x.com/TylerG_Capital/status/2094117526942761382
A trader spent four weeks with Claude fine-tuning the models he trades so the alerts do the work. Focusing on NQ futures and native Bitcoin for crypto, his entries are now entirely mechanical, he just manages positions off the alerts instead of staring at charts. His framing is that the point was eliminating his own emotion and over-analysis, the thing that hurt him for years, and that coding these indicators to the finest detail is how he took discretion out of the loop.
@DaviddDotTech [Claude Code]
https://x.com/DaviddDotTech/status/2093990899315290230
A concrete recipe for swapping the engine: he runs GLM 5.2 inside Claude Code by pointing the config at a GLM key, then wires in a backtesting MCP server and tells it to loop every five minutes improving a BTCUSDT strategy until it finds something profitable. His result came back at 230% net profit, 477 trades, 1% max drawdown, profit factor 5.58 on the 5-minute timeframe, at roughly a quarter of Claude's cost. He notes it's slower than Claude but brilliant at optimizing, and to forward-test before any real money.
@Damir_Akaza [Claude Code]
https://x.com/Damir_Akaza/status/2094086316195516581
Dropped a single prompt into a four-agent group chat, went to grab a coffee, and the bots split the work, ran local scripts over a Mac Mini via Tailscale, generated 50 creative variations while fact-checking offers in parallel, and pushed 9 approved ads live into Meta Ads Manager, 40 minutes, zero manual setup. He's honest about the downsides too: the agents still fall short of Claude Code on copy and hallucinate, but the autonomous media-buying pipeline is real.
@ChShersh [Claude Code]
https://x.com/ChShersh/status/2093951465215713466
A small but genuine learning case. He implemented std::unique_ptr from scratch, then asked Claude Code to review his implementation. He calls it an amazing learning experience, surfacing subtle nuances he hadn't thought about. Not a build-my-app story, an agent used as a demanding code reviewer that teaches you the corners of your own work.
@connect24h [Claude Code]
https://x.com/connect24h/status/2093993440904380568
A workflow running at GMO Pepabo that reframes how you hand work to an agent. When a reviewer marks a nit as "later" or "separate PR," you just slap a label on the PR. Overnight, GitHub Actions and Claude Code pick it up and run implementation to a develop-bound PR to auto-merge. The clever part is you never rewrite an issue: the context already lives in the PR review thread, so the human effort is one label. Riskier changes like infrastructure and DB schema stay excluded for human review.
@light940 [Claude Code]
https://x.com/light940/status/2094203259879850454
Wantedly is pushing AI adoption with the goal of "an embedded data analyst for every employee." What's worth copying is the structure: they split data-access design from platform development, hand the dbt implementation to Claude Code, and keep humans focused on requirements and review. It's a concrete version of the AI-native pattern, agent owns execution, people own defining the problem and approving the result.
@cboyack [OpenClaw]
https://x.com/cboyack/status/2094173933620847046
A long list of real personal-automation workflows he ran through OpenClaw (now moving to a newer agent, but the jobs are the point): turns on his Control4 bedroom lights to wake up and turns them back on if he snoozes, monitors flights and pre-conditions his Tesla's climate so the car is comfortable on landing, uses Canva to make podcast thumbnails and SRT files, surfaces three meal options he hasn't had lately and handles the DoorDash order around his calendar, and two weeks before travel checks his CRMs for donors or VIPs in the area to connect with.
@Lynx_0C [OpenClaw]
https://x.com/Lynx_0C/status/2094074624334336239
Took Miloco, Xiaomi's OpenClaw plugin for integrating cameras for environmental awareness, and did his own second-layer development on it: added support for OpenAI-standard endpoints for inference (originally cloud-only Xiaomi models), RTSP cameras (originally Xiaomi-only), and Home Assistant control. His note is that this kind of local perception clearly belongs on locally-run models, and he's opened it for testing.
🗣 User Voice
User Voice
The quota change is the loudest thing in the feed, and it's not really about the number, it's about being talked to like you can't do arithmetic. @ai_for_success called the "permanent 25% increase" post a clown move that actually works out to a 17% reduction (https://x.com/ai_for_success/status/2093911662277558367), and dozens echoed the comparison: OpenAI reset Codex allowances the same week and fixed quota-eating bugs while Anthropic dressed a cut as a raise. What users want is honesty and a live meter, not a math puzzle.
The mental model has shifted from "which model" to "which harness." @realchendahuang laid it out cleanly: the model decides what to do, the harness actually reads files, writes to disk, runs shells, manages context, so the real question is no longer whether DeepSeek or GLM is strong, it's which harness makes that model strongest (https://x.com/realchendahuang/status/2093889637936926759). Users want the model to be a swappable part they can route around.
Memory that survives across sessions and across agents is a recurring ask. @heydittoai captured the pain: switching between three agents to finish one task means you become the memory, re-explaining the stack and last Tuesday's decisions each time (https://x.com/heydittoai/status/2094167527047999607). People want one shared context under whatever model they move to next, not per-vendor memory silos.
Token waste is the other cost complaint, and it's structural, not about the model. @MrAhmadAwais benched shell tools across harnesses and found roughly 306k of every 1M shell-driven tokens is removable waste, mostly from polling and re-reading the same logs (https://x.com/MrAhmadAwais/status/2094165837913747810). The wish is for harnesses that stop scaling garbage as fast as they scale work.
Once you run more than a handful of agents, the bottleneck becomes your own attention. @jjacky, after using half a dozen agent interfaces, said no one has solved the UI for working with many parallel sessions, and he still gets lost past four (https://x.com/jjacky/status/2093924394632401126). The unsolved problem isn't the agents, it's the cockpit for supervising them.
The quota change is the loudest thing in the feed, and it's not really about the number, it's about being talked to like you can't do arithmetic. @ai_for_success called the "permanent 25% increase" post a clown move that actually works out to a 17% reduction (https://x.com/ai_for_success/status/2093911662277558367), and dozens echoed the comparison: OpenAI reset Codex allowances the same week and fixed quota-eating bugs while Anthropic dressed a cut as a raise. What users want is honesty and a live meter, not a math puzzle.
The mental model has shifted from "which model" to "which harness." @realchendahuang laid it out cleanly: the model decides what to do, the harness actually reads files, writes to disk, runs shells, manages context, so the real question is no longer whether DeepSeek or GLM is strong, it's which harness makes that model strongest (https://x.com/realchendahuang/status/2093889637936926759). Users want the model to be a swappable part they can route around.
Memory that survives across sessions and across agents is a recurring ask. @heydittoai captured the pain: switching between three agents to finish one task means you become the memory, re-explaining the stack and last Tuesday's decisions each time (https://x.com/heydittoai/status/2094167527047999607). People want one shared context under whatever model they move to next, not per-vendor memory silos.
Token waste is the other cost complaint, and it's structural, not about the model. @MrAhmadAwais benched shell tools across harnesses and found roughly 306k of every 1M shell-driven tokens is removable waste, mostly from polling and re-reading the same logs (https://x.com/MrAhmadAwais/status/2094165837913747810). The wish is for harnesses that stop scaling garbage as fast as they scale work.
Once you run more than a handful of agents, the bottleneck becomes your own attention. @jjacky, after using half a dozen agent interfaces, said no one has solved the UI for working with many parallel sessions, and he still gets lost past four (https://x.com/jjacky/status/2093924394632401126). The unsolved problem isn't the agents, it's the cockpit for supervising them.
📡 Eco Products Radar
Eco Products Radar
Products and tools mentioned three or more times across today's posts:
Codex, the constant comparison point, with users noting its quota resets and desktop app while debating CLI parity with Claude Code.
Grok Bot, xAI's new agent, repeatedly framed as the "easy-mode" harness that runs Claude Code, Codex, and Cursor underneath it.
Cursor, in the news over the OpenAI model-access cutoff and its SpaceX acquisition.
Hermes, Nous Research's self-improving agent, paired constantly with OpenClaw in multi-agent stacks.
OpenClaw, its 2.0 release nearing amid a wave of "does anyone still use it" posts.
Pi, the minimal harness several power users run as their core with a handful of extensions.
GLM-5.3 and DeepSeek, the open-weight models people route into Claude Code to cut cost.
Obsidian, the vault at the center of the Karpathy-style second-brain workflow.
Antigravity and Kimi, recurring alternates in the coding-agent rotation.
Products and tools mentioned three or more times across today's posts:
Codex, the constant comparison point, with users noting its quota resets and desktop app while debating CLI parity with Claude Code.
Grok Bot, xAI's new agent, repeatedly framed as the "easy-mode" harness that runs Claude Code, Codex, and Cursor underneath it.
Cursor, in the news over the OpenAI model-access cutoff and its SpaceX acquisition.
Hermes, Nous Research's self-improving agent, paired constantly with OpenClaw in multi-agent stacks.
OpenClaw, its 2.0 release nearing amid a wave of "does anyone still use it" posts.
Pi, the minimal harness several power users run as their core with a handful of extensions.
GLM-5.3 and DeepSeek, the open-weight models people route into Claude Code to cut cost.
Obsidian, the vault at the center of the Karpathy-style second-brain workflow.
Antigravity and Kimi, recurring alternates in the coding-agent rotation.
Comments