Super User Daily: August 20, 2026
The loudest signal today is memory: five separate builders shipped or pitched persistence layers on the same day, because everyone is tired of their agent forgetting everything at session close. Meanwhile the most interesting use cases keep drifting away from code: a Japanese trader spent a year teaching Claude Code his charting rules and documented every way the model cut corners, a 3D-printing hobbyist wired an OpenClaw-style CAD agent, and a legal services company put 50+ agents into production built by non-engineers. On the OpenClaw side the mood is turning: heavy users are publicly drifting to Grok Bot, a broken Telegram integration sat in a beta for three days with zero issues filed, and one user asked the question out loud: why did everyone stop talking about it at the dinner table?
@vikasg_rac [Claude Code]
https://x.com/vikasg_rac/status/2089548722573398195
Works in hours-long Claude Code sessions and got tired of scrolling one continuous terminal stream where machine output buries the actual conversation. So he built claude-reader with Claude Code itself: a tiny terminal app that tails the transcript Claude Code already saves in ~/.claude/projects/, filters it down to prompts and replies, and renders them as markdown in a side pane, refreshing twice a second. No hooks, no plugins, no server, works over ssh, and UI changes to Claude Code cannot break it because it only reads a file. Before release he had a different model review the code, which caught four real crashers that Claude then fixed and covered with tests.
@Xudong07452910 [Claude Code]
https://x.com/Xudong07452910/status/2089512853682479451
ABC Legal, an 1,100-person legal services company, handed Claude Code to 15 non-engineers from finance, marketing and ops. One week later all 15 had working agents; about a month later the company was running 50+ agents in production. The scaling trick is treating agents as code: prompts, tools, memory and schedules all live in Git, every change goes through a PR, so versioning, review, rollback and audit come for free. They even built a self-improvement loop where a harvester agent collects Slack emoji feedback and a tuner agent proposes prompt changes as PRs for humans to approve. Part of the workforce cost dropped about 50 percent while total AI cost went down as agent count grew.
@KaiKai90000 [Claude Code]
https://x.com/KaiKai90000/status/2089848264757837872
A Japanese retail investor documented a full year of trying to make AI analyze stock charts, and the honest failure log is the valuable part. Chat-based approaches gave unstable answers and investment-advice guardrails returned textbook lectures instead of signals; the unlock was switching from asking AI for answers to assigning AI work, rebuilding the system in Claude Code. His hardest-won lesson: when he fed a whole trading methodology at once, the model said done while silently dropping rules, and only admitted the gaps when challenged. His fix is now procedure: feed material block by block, verify each block's output, never trust a bare done. The prototype is running in live demo mode and the new line launched August 10 inherits all of it.
@withkate_ [Claude Code]
https://x.com/withkate_/status/2089831706807484508
Her go-to screenshot extension GoFullPage, 11M+ downloads, vanished from the Chrome store over a copyright dispute. Instead of waiting she rebuilt it with Claude Code in under 30 minutes: scrolls the page, stitches one image, auto-saves as png, jpg or pdf, and even handles sticky navs. It is fully local and open source, loaded via developer mode. Her own summary of the shift is the story: the first reaction to losing a tool is no longer frustration, it is I'll just build it myself.
@anirudh_twt [OpenClaw]
https://x.com/anirudh_twt/status/2089606129081045306
Runs OpenClaw plus GBrain with a custom layer that lets him use his Claude Max subscription in the cloud, with Claude Code running in daemon mode as the always-on brain. The layer persists context across sessions, spins off Codex for coding and Sonnet for writing, and at the end of every week the system reviews its own work, harvests repeated patterns into skills, and scores every agent on where it excelled and failed. He describes it as managing very smart coworkers where you choose between micromanaging and delegating broad problems. The setup is 90 days old and cost $220 in software plus $1,150 in cloud bills.
@Lucastcaraujo34 [Claude Code]
https://x.com/Lucastcaraujo34/status/2089837362243141884
Runs weeks-long loops across three Codex Pro accounts and two Claude Max accounts, and most of the wall-clock time is spent waiting for rate limits to reset. His architecture: Hermes generates an orchestration file that works like a procedure with forms, Claude Code running Fable 5 reads it and orchestrates subagents that must fill the forms in. The next step is a curator agent that aggregates all the review-loop transcripts into one idealized first-try-perfect transcript per task, which becomes training data for local models. The goal is loops that run for days without cost anxiety.
@viticci [OpenClaw]
https://x.com/viticci/status/2089708483444625660
Federico Viticci calls Open Minis the most interesting indie iOS app of his year: effectively a mini sandboxed OpenClaw that runs natively on iPhone. The developer turned dozens of Apple frameworks including Photos and HealthKit into CLIs inside the app, shipped a built-in Alpine Linux terminal to call them, and gave the agent its own Files workspace. It runs any OpenAI-compatible model, mounts an Obsidian vault, and has built-in browser use via WebKit. Viticci recreated a self-improvement loop inside it and triggers curation sessions via Shortcuts automations, which is about as power-user as iOS gets today.
@laoyingkhq [OpenClaw]
https://x.com/laoyingkhq/status/2089614654842781913
A trader reportedly opened a fresh Polymarket account with $1,000, wired up an arbitrage bot on OpenClaw, and turned it into $5,500 within 48 hours. The post links the setup publicly rather than selling a course, which is why it spread. Worth flagging: prediction-market arbitrage is exactly the kind of loop where agents outrun humans on speed, and also exactly the kind of claim that deserves skepticism until you have seen the fills. Either way, agent-driven market operations keep showing up as OpenClaw's most lucrative reported use case.
@ChuckCook [OpenClaw]
https://x.com/ChuckCook/status/2089736459905368229
Has run OpenClaw, Hermes and now Grok Bot side by side, and his current project is the interesting part: an agent that handles CAD modeling for his Prusa 3D printer. He wired Build123d into a custom skill that interviews him for the part's details and then renders the result in an OCP CAD viewer, and one-shotted a replacement coffee scoop design. This is the physical-output end of agent use: describe an object, get a printable model. His verdict on the harness war so far is that Cursor's polish is winning him over from the setups he built himself.
@startupideaspod [Claude Code]
https://x.com/startupideaspod/status/2089761351707316523
Gave Claude Code eyes and made post-build verification a standard pass: run the app, open the page, adopt a persona matching the target customer, and audit the flow like a first-time visitor. In his test the agent clicked the waitlist CTA with an empty form, checked the backend record, then swept console and network for errors. He caps it at one focused pass on the single highest-impact issue, which in his case surfaced a missing expectation-setting microcopy at the CTA. The principle: a build is not a check, and a page can load and still be confusing.
@Syuiro_2 [Claude Code]
https://x.com/Syuiro_2/status/2089547330622214348
Systematized a method for analyzing data you must never show an AI: use Claude Code's coding ability while keeping confidential data strictly local. The model writes and iterates the analysis scripts, the scripts run against the sensitive data, and only schemas and aggregate results ever reach the model. For anyone in finance, healthcare or legal who has been blocked from using agents by compliance, this borrow-the-hands-not-the-eyes pattern is the workaround, and the post organizing it resonated widely in the Japanese community.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2089509583312339450
Built a migration prompt that makes Claude Code or Codex package an entire agent setup for transfer into Grok Bot swarms: every skill file with summaries, memory condensed into one profile document, every recurring workflow with trigger and schedule, plus a suggested breakdown of workflows into team roles like CFO and research assistant. Whatever you think of the destination, the pattern matters: your agent configuration is now a portable asset, and the exit cost from any harness is one well-written export prompt.
@charliejhills [Claude Code]
https://x.com/charliejhills/status/2089744513376174571
Shipped three free skills that attack Claude Code's readability problem. /show-me renders whatever Claude built as a real page and brings the browser forward instead of describing it in text; /answer-first forces the answer onto line one with the next step at the bottom; /diagram-design gives Claude 27 diagram types so structures arrive drawn. He also shares a CLAUDE.md rule making every viewable deliverable a self-contained HTML file that Claude must render and verify before calling done. Two months of daily use, and for visual thinkers it changes the tool's character.
@obie [Claude Code]
https://x.com/obie/status/2089831991860719715
Obie Fernandez's fix for the blown-cache problem: when you return to Claude Code more than an hour after the last turn, don't resume and eat the cost of re-priming the full context. Exit, copy the session id, start a fresh session and say continue where session X left off. The agent gathers exactly as much history as it needs from the old transcript and refers back on demand. He suggests this should be built-in behavior, and given how many people idle their sessions, the token savings are real.
@poetengineer__ [Claude Code]
https://x.com/poetengineer__/status/2089772616533868615
Visualized one week of Claude Code work as a growing structure built up day by day in chronological order, where hovering over a project spreads it into per-file changes. It is a small art-engineering piece, but it points somewhere practical: agent transcripts are rich structured data that almost nobody looks at, and rendering them makes weekly output legible in a way commit logs never are.
@notEgoyard [Claude Code]
https://x.com/notEgoyard/status/2089729592118431913
Built a programmatic motion design studio inside Claude Code that eliminates manual keyframing: Claude reads Figma design tokens and converts specs into responsive vector graphics, Remotion handles frame-by-frame animation with spring physics, Three.js supplies shader backgrounds, and a ComfyUI API layer generates style-consistent assets. The control mechanism is a brand kit json holding grid, type scales, motion constants and palette, which turns Claude into an art director that keeps every frame on-brand. Brief in, broadcast-grade 60fps video out.
@aacle_ [Claude Code]
https://x.com/aacle_/status/2089737234358702401
A bug bounty hunter's take on restarting from zero: skip collecting 100 tools, build one solid workflow and put Claude Code inside it from day one. The thread walks through recon, triage and report drafting as agent-assisted stages rather than tool checklists. Security research is quietly becoming one of Claude Code's strongest non-product niches, and the framing here, workflow over arsenal, is why.
@HaoliYin [Claude Code]
https://x.com/HaoliYin/status/2089729055528751411
Research result with a sharp diagnosis: general coding agents lack research taste, making small tweaks and getting stuck in local optima. Their system DataSmith automates data research by inspecting failures, searching wider, and making better research bets, beating Claude Code by 5.1 points on average. The gap it exposes matters more than the tool: harnesses tuned for shipping code behave badly on open-ended experimentation, and specialized research harnesses are becoming their own category.
@aigclink [Claude Code]
https://x.com/aigclink/status/2089530764706861145
PenguinHarness, from the LlamaFactory author Zheng Yaowei, is an agent that builds agents: one sentence generates a full agent application with scaffolding, code and docs, and the demo RAG app cost 0.2 yuan of tokens. It keeps the toolset deliberately small, claims top accuracy on data analysis comparisons at 1/70th the cost of Claude Code, and ships self-evolution skills where the agent benchmarks, evaluates and optimizes itself with rollback snapshots each round. Deeply adapted to Chinese open models but speaks any OpenAI-compatible endpoint.
@janwilmake [Claude Code]
https://x.com/janwilmake/status/2089738039748653521
Built agent-codemode, then discovered mcporter had beaten him to the idea, and his response is a small masterclass in open source manners: the one real difference is that his 300-line version reads Claude Code's keychain tokens directly instead of keeping its own vault, killing the auth step entirely. He is offering it as a PR to mcporter rather than running a second thing. The keychain trick is the takeaway: your installed Claude Code is already an auth provider other tools can piggyback on.
@DanKornas [Claude Code]
https://x.com/DanKornas/status/2089847315133546669
hipocampus is a file-based memory harness for Claude Code, OpenCode and OpenClaw built on a three-tier design: an always-loaded ROOT.md topic index of about 3K tokens, on-demand files, and searchable history compacted hierarchically from raw logs into daily, weekly and monthly summaries. Recall combines index triage, manifest selection and BM25 plus vector search. One command installs it with platform detection. MIT licensed, and one of five memory plays that shipped or resurfaced on the same day.
@HelloVyom [Claude Code]
https://x.com/HelloVyom/status/2089714975636144354
Graft attacks the team version of agent amnesia: everything your agent learns about a repo evaporates at session close, and your teammate's agent relearns it from scratch tomorrow. It generates a markdown repo map, saves it into the project, updates only when code changes, and travels through git so the whole team's agents share it. Runs locally with no accounts or keys, open source under MIT. The pitch line lands: a new hire reads the docs once, your agent has not read anything in months.
@emerginghope_ [Claude Code]
https://x.com/emerginghope_/status/2089841774055301576
A sharp critique of the synced-memory trend: two agents can remember the same thing and still leave you with two versions of the company, like Claude updating the launch plan while Codex acts on yesterday's decision. His OrgX is a shared work-state layer rather than a bigger context pile: which initiative is active, which decision replaced which, what needs approval, what artifact came back. Setup is a wizard that wires local clients into a hosted MCP. His closing distinction is the memorable part: a local vault helps the agent remember you, OrgX is for when agents start remembering different companies.
@indelibleBSV [Claude Code]
https://x.com/indelibleBSV/status/2089838144279466348
A founder confession with a lesson about MCP: for months he marketed Indelible as permanent AI memory inside Claude Code and VS Code, then discovered mid-onboarding that his user runs Cursor, and realized it never mattered. Indelible is an MCP server, so anything speaking the protocol gets the memory layer. He built for the standard and marketed the smallest room in the house. Claude Code keeps the deepest integration with automatic zero-click saving, but the addressable market was always every MCP client.
@arjunkshah21 [OpenClaw]
https://x.com/arjunkshah21/status/2089839743232303274
A 14-year-old's open-source prompt compression tool crossed 500 users and 1k+ weekly npm downloads, and has processed over 90 million tokens since launch. It analyzes context and query, keeps what matters, and cuts about 65 percent of the input while retaining over 98 percent of critical content. It ships as an API and an MCP, plugging into Hermes, OpenClaw, Claude Code, Cursor, Codex and the rest. Context costs are now a big enough pain that a teenager can build real adoption attacking them.
@amazing129 [OpenClaw]
https://x.com/amazing129/status/2089627937771585774
A loyal OpenClaw user's health report doubles as a warning sign: a three-day-old beta image shipped with Telegram completely broken, and nobody had filed an issue or PR, unthinkable during the peak months. He fixed it himself and submitted the patch, then found the official Railway deploy templates stale and unmaintained, so he published his own updated repo. He remains a believer in the design philosophy while documenting exactly what a community exodus looks like from the inside: the crowd that files issues has moved on.
@agentavow [OpenClaw]
https://x.com/agentavow/status/2089768806906269824
A security shop scanned 231 OpenClaw skills and found 14,350 issues, with 32 percent of skills scoring an F. The supply chain behind agent skills is the story: every skill is third-party code with real permissions, and the ecosystem grew far faster than any review culture. Combined with this week's warnings about tampered CLAUDE.md files planted to leak API keys, the pattern is clear, agent extensions are the new npm typosquatting surface, arriving before the tooling to defend it.
@masonltompkins [OpenClaw]
https://x.com/masonltompkins/status/2089843299142275225
Building AYGENT, an agent harness like OpenClaw for people who hate terminals, and doing it in public with a twist: he develops the app from inside the app, and posted an unedited screen capture, typos included, of what that dogfooding loop looks like. The niche is real, every current harness assumes terminal fluency, and the non-terminal majority is unserved. The build-with-itself loop is both his QA and his demo.
@Yumeira9 [Claude Code]
https://x.com/Yumeira9/status/2089838089976078413
Built an AI secretary with a second brain: a full Obsidian vault holding bad-news routing rules, briefing schemas, a no-surprise protocol and five governed profiles. Output is validated against strict schemas and the agent halts before any action until approved. It runs on Claude Code and Codex and she talks to it through her phone. This is the governance-heavy end of personal agents, where the interesting engineering is in the rules about when the agent must stop.
@DawsonSchrader [Claude Code]
https://x.com/DawsonSchrader/status/2089839073376542794
Designed and built his own terminal multiplexer as an Obsidian plugin: tmux sessions that survive Obsidian closing, plus a management sidebar, all scoped to his very specific business workflow. Then herdr arrived and he is rewiring the sidebar to open spaces there instead. The kicker was portability: none of his memory was stuck in Claude Code, so Grok 4.6 picked up his business, his clients and his custom vault immediately. Personal infrastructure, disposable harnesses.
@gullett_d [Claude Code]
https://x.com/gullett_d/status/2089772963172135289
Expected Microsoft Defender to find Claude Code and maybe Cursor on his lab machines. It found eight AI agents across two devices, his phrase: my endpoints were running a small bed-and-breakfast. He dug into the inventory, MCP servers, AgentsInfo, licensing and runtime protection and wrote it up. Enterprise security tooling is starting to see the agent sprawl that employees installed one experiment at a time, and most orgs have no idea what their own count is.
@kv1nsiii [Claude Code]
https://x.com/kv1nsiii/status/2089766030516383885
The story of a 70-line markdown file with nearly 200K combined stars across mirrors: it teaches Claude nothing new, it just says what not to do. Don't overengineer beyond the existing complexity, don't touch code you weren't asked to touch, don't make silent assumptions, don't invent libraries. Built by a developer the day after Karpathy publicly named the failure patterns. The security footnote matters: tampered copies exist planted specifically to leak API keys, so pull it only from the official repo.
@v4rvl [Claude Code]
https://x.com/v4rvl/status/2089843954980176033
Made Unslop, a local skill plus hooks that force Claude Code, Codex and Grok to write in clear English, covering the main thread, subagents and their outputs. The hook approach is the point: politely asking a model to drop the slop decays over a long session, while a hook re-enforces it on every turn. Small tool, real quality-of-life gain for anyone drowning in AI prose.
@fujibee [Claude Code]
https://x.com/fujibee/status/2089847397648351433
Runs large fleets of Claude Code instances and hit the wall everyone scaling agents hits: each instance is too heavy to multiply. His read of Vercel's new fx harness is the practical one, a single native Zig binary, 10 microsecond cold start, 6.3MiB with single-digit MB baseline memory, Apache licensed with no telemetry. His planned formation: keep heavyweight Claude Code as the leader and line up single-digit-MB fx workers underneath. Harness weight is becoming a real architecture variable.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2089691030023250414
Ran a two-tool stack for six months, Claude Code for building and Hermes for daily ops, and thought he was covered. Grok Bot exposed a third job he had never assigned: the boring 24/7 watching, community questions at 3am, testing the app while he is at the gym, researching clients before he wakes. His conclusion generalizes: that job needs something always awake, not something maximally smart, and the division of labor between smart agents and vigilant agents is becoming its own design decision.
@Granite0x [Claude Code]
https://x.com/Granite0x/status/2089496993299058710
The monthly ritual of losing months of Claude Code work to an expired subscription turned out to be an illusion: the chats were on disk the whole time, in a format only the app could read. Someone open-sourced ai-data-extractor, one pip install that cracks the local stores of six assistants, Claude Code, Cursor, Windsurf, Aider, Cline, Roo Code, and exports prompts, responses, diffs, tool calls and timestamps as plain files. No account, no re-subscribe. Your transcripts were always yours; now you can actually open the door.
@ClaudeCode_love [Claude Code]
https://x.com/ClaudeCode_love/status/2089683227690869131
A Japanese user distilled Anthropic's official cost-reduction guidance into four moves and reported which one actually paid: set the quality bar first so cheaper runs have a pass line, stop resending the same preamble every call, don't front-load all reference material but let the agent fetch on demand, and pre-specify output shape to shorten responses. The second one, prompt reuse via caching, roughly halved his per-task cost by itself. Boring advice, measured result.
@bprintco [OpenClaw]
https://x.com/bprintco/status/2089573163579384221
Already owned a Mac mini before OpenClaw existed, then dedicated the machine to it and calls agentic video editing on that box phenomenal. Three separate users today described the same pattern, a Mac mini as a standing agent appliance, running loops and local models around the clock. The always-on local agent box is quietly becoming a hardware category, assembled from consumer parts before any vendor names it.
🗣 User Voice
User Voice
1. The pricing ceasefire is temporary and everyone knows it. Anthropic extended the 50 percent higher Claude Code limits through August 31 while warning capacity may be tight, and users are already planning for the cliff: detailed cost breakdowns show default subagent costs nearly tripled across two model generations, and the recommended defenses are pinning cheaper subagents or routing bulk work elsewhere (@nauczymycieAI, @danielcberk, @nakakei6439).
2. A vocal cluster of Max subscribers reports the same degradation: long thinking, then slow, off-track, low-quality output that feels unusable for real work, with no clear explanation (@0xQuantic). Others describe Claude Code as slower and less sharp than before and are openly auditioning Codex (@asumimirainet), and one engineer notes CLAUDE.md instructions get ignored exactly when a model regression lands (@lavishG10).
3. OpenClaw's community health is the quiet story of the day: nobody talks about it at the dinner table anymore and no one agreed to stop (@CtrlAltDwayne), a broken Telegram integration sat unreported in a beta for days (@amazing129), and switchers cite way too buggy as the reason (@startupstella). The project's design ideas remain influential; its maintenance energy is visibly draining.
4. Memory is the most demanded missing layer, and users want it portable: five separate persistence tools surfaced in one day (@HelloVyom, @DanKornas, @emerginghope_, @indelibleBSV, @anirudh_twt), and the sharpest user critique is that synced memory without shared work-state just produces agents that confidently remember different realities (@emerginghope_).
5. Users want agents that verify, not just generate: the give-it-eyes verification pass (@startupideaspod), block-by-block feeding with per-block checks because the model will silently skip rules (@KaiKai90000), and cross-model code review before release (@vikasg_rac) all converge on the same demand, a built-in adversarial check on done.
1. The pricing ceasefire is temporary and everyone knows it. Anthropic extended the 50 percent higher Claude Code limits through August 31 while warning capacity may be tight, and users are already planning for the cliff: detailed cost breakdowns show default subagent costs nearly tripled across two model generations, and the recommended defenses are pinning cheaper subagents or routing bulk work elsewhere (@nauczymycieAI, @danielcberk, @nakakei6439).
2. A vocal cluster of Max subscribers reports the same degradation: long thinking, then slow, off-track, low-quality output that feels unusable for real work, with no clear explanation (@0xQuantic). Others describe Claude Code as slower and less sharp than before and are openly auditioning Codex (@asumimirainet), and one engineer notes CLAUDE.md instructions get ignored exactly when a model regression lands (@lavishG10).
3. OpenClaw's community health is the quiet story of the day: nobody talks about it at the dinner table anymore and no one agreed to stop (@CtrlAltDwayne), a broken Telegram integration sat unreported in a beta for days (@amazing129), and switchers cite way too buggy as the reason (@startupstella). The project's design ideas remain influential; its maintenance energy is visibly draining.
4. Memory is the most demanded missing layer, and users want it portable: five separate persistence tools surfaced in one day (@HelloVyom, @DanKornas, @emerginghope_, @indelibleBSV, @anirudh_twt), and the sharpest user critique is that synced memory without shared work-state just produces agents that confidently remember different realities (@emerginghope_).
5. Users want agents that verify, not just generate: the give-it-eyes verification pass (@startupideaspod), block-by-block feeding with per-block checks because the model will silently skip rules (@KaiKai90000), and cross-model code review before release (@vikasg_rac) all converge on the same demand, a built-in adversarial check on done.
📡 Eco Products Radar
Eco Products Radar
DeepSeek Harness (dsh) - the open-source harness kept dominating conversation: 148K stars in 4 days, everything-is-a-plugin architecture, subagent bridges to Claude Code and Codex.
Grok Bot (@bot) - the gravitational pull of the day; multiple OpenClaw and Claude Code users describe migrating daily-ops workloads to it.
Hermes Agent - repeatedly cited as the ops layer beside Claude Code, and as the orchestration-file generator in multi-account setups.
Codex - the default alternative users audition whenever Claude Code frustrates them; also the co-runner in most dual-agent setups.
Cursor - polish repeatedly praised by harness veterans; now shipping Grok Bot integration.
OpenCode - recurring as the open CLI alternative in multi-agent fleets and router setups.
GBrain - agent memory and skills layer appearing in cloud setups, bootstrap flows, and autoresearch indexing.
Model routers (claude-code-router, 9router, CC Switch, Codex Subscription Router) - an entire category now: routing by privacy, cost, quota and account, under the same workflow.
Mac mini - the unofficial agent appliance: three users independently described dedicating one to always-on agents and local models.
Obsidian - the vault keeps showing up as the human-readable memory substrate agents share.
DeepSeek Harness (dsh) - the open-source harness kept dominating conversation: 148K stars in 4 days, everything-is-a-plugin architecture, subagent bridges to Claude Code and Codex.
Grok Bot (@bot) - the gravitational pull of the day; multiple OpenClaw and Claude Code users describe migrating daily-ops workloads to it.
Hermes Agent - repeatedly cited as the ops layer beside Claude Code, and as the orchestration-file generator in multi-account setups.
Codex - the default alternative users audition whenever Claude Code frustrates them; also the co-runner in most dual-agent setups.
Cursor - polish repeatedly praised by harness veterans; now shipping Grok Bot integration.
OpenCode - recurring as the open CLI alternative in multi-agent fleets and router setups.
GBrain - agent memory and skills layer appearing in cloud setups, bootstrap flows, and autoresearch indexing.
Model routers (claude-code-router, 9router, CC Switch, Codex Subscription Router) - an entire category now: routing by privacy, cost, quota and account, under the same workflow.
Mac mini - the unofficial agent appliance: three users independently described dedicating one to always-on agents and local models.
Obsidian - the vault keeps showing up as the human-readable memory substrate agents share.
Comments