No-Code Agentist

The Hermes Company Masterclass, part 1 of 2. What real operators run on Hermes, limits stated first, then your week as routines, a charter, an org of bots with human gates, a chief of staff, and production lines. Free: https://t.co/fJgpc03O7y https://t.co/W2xNgWQIfW

GROK BOT Sources: official https://t.co/2EN1BbJ794 + https://t.co/0OVb26msAp + your Grok Bot bookmark folder (948 posts) Updated: 2026-09-19 NEW TODAY - Folder 948 posts (15 new vs Sep 18) - Official (bot, Sep 18): Grok Bot can now send you voice notes - Testers (techdevnotes, Sep 18): Grok Bot now has Voice notes (quotes official bot) - SpaceXAI notes (blankspeaker, Sep 18): Version 0.57.0 — steer away from group chats; see in-progress tasks + text; see your tasks list - SpaceXAI notes (blankspeaker, Sep 18): Version 0.57.1 — bugfixes - SpaceXAI notes (blankspeaker, Sep 18): Testing Expert-mode chat queries routed through the Grok Bot harness on https://t.co/UvO8aC6o4n (multi-step search/compare/timeline vs single-pass) - Grok Bot team (ericzakariasson, Sep 18): Recap of what shipped the last couple weeks — marketplace + bot templates (and more in thread) - Testers (ab_workss, Sep 18): Grok bot phone widget — multi-Bot overview UI or single-Bot focus UI - Testers (techdevnotes, Sep 18): Marketplace adds Runway, Zernio, Migma, Listen Labs, Perspective AI, Read AI, HTML/CSS to Image, Jam, Nitrosend, Paybox, ReadMe, Adspirer, Browserless, Modem, Orthogonal plugins - Testers (liam_fallen, Sep 18): Safer-online Bot — scans email for accounts, password manager, devices/sessions, forgotten accounts, 2FA/passkeys, app access, recovery (never asks for your passwords) - Testers (GrokBotRadar, Sep 18): Highlight — Liam Fallen privacy Bot hunts exposed personal info, files removals, follows up if ignored - Testers (pengzheng_, Sep 17): NotebookLM-style podcast Bot — send a link to a Podcast bot; listen-ready episode + sideload to podcast app; opens sites NotebookLM can’t - Testers (0xRafy, Sep 18): Nightly Engineer Bot works while you sleep — researches codebase nightly, finds cleanup opportunities - Testers (XFreeze, Sep 19): SpaceXAI CFO Bret Johnsen — Grok Bot growing faster than anything; cool innovative tool for efficiency - Testers (bloggersarvesh, Sep 18): 20 SEO prompts pack for Grok Bot (open Bot → paste prompts → outrank competitors) - Testers (doganuraldesign, Sep 18): 3-year OneChat AI dream now real with Grok Bot WHAT IT IS - A Bot is a named, persistent teammate, not a one-off chat - You message it like a coworker and it finishes work in real apps - All of your Bots share one cloud computer (browser, files, terminal) - Each Bot has its own screen on that shared computer - Work keeps going when your laptop or the app is closed - It only comes back when it needs approval or a human step - Do not confuse Grok Bot with Digital Optimus (real-time video human emulation) - Testers (joehansen): do not confuse Grok Bot with an X Chat Agent (an X account you add to a group chat). Grok Bot is the machine that does the work - Official Elon: Your Bot will identify issues to resolve, notify you when they are fixed, and continue on with your project ACCESS AND INSTALL - Official bot (Aug 26): All SuperGrok and Cursor Pro subscribers now have access - Testers (mattyp): Grok Bot is free to try; every Grok & Cursor plan includes Grok Bot - Testers quoting SpaceXAI (cb_doge): SuperGrok, SuperGrok Plus, SuperGrok Heavy, Cursor Pro, Cursor Pro+, Cursor Ultra, Cursor Teams Standard and Premium - Cursor CEO Michael Truell: everyone with a standard Grok or Cursor subscription - Testers: included at no extra cost with those plans; circulating plan prices $20/month Cursor Pro and $25/month SuperGrok (nextbigfuture) - Official bot + Elon: weekly / free usage limits reset for all Grok Bot users - Sign in with your Cursor account (that account owns plan and usage) - Official desktop download: - Also listed at - Official desktop is macOS and Windows (Apple silicon or Intel; Windows x64 or Arm64) - Testers (mattyp): Linux is now supported (official docs may lag) - iPhone: iOS 18+, App Store app "Grok Bot" - Official (bot, Sep 2): Grok Bot is now available on Android - Testers (mark_k): install from Play Store; full Grok Bot experience on Android - Play Store: - Official docs may still lag on Android; Testers (mattyp, Sep 4): Grok Bot is now available for iPad - SuperGrok Heavy users can choose Get access with SuperGrok Heavy, then Link Grok Account - Linking SuperGrok Heavy can unlock free Cursor Ultra (one Grok account to one Cursor account) - Testers (unicodef1wn, Sep 10): plans don’t stack — linking SuperGrok to a Cursor plan adds zero usage; upgrade within Cursor or turn on on-demand if you need more - Legacy Privacy Mode blocks Grok Bot; switch to a supported Cursor data setting first - Grok Bot checks for updates automatically; Check for Updates is in Settings → Beta - Official Elon (Sep 4): Grok Bot Enterprise is now available - Testers (mark_k, Sep 3): Version 0.36.0 is now live — bugfixes and improvements - SpaceXAI notes (blankspeaker, Sep 9): Version 0.47.0 — auto lock team review; auto review team rules - Official Elon (Sep 17): Grok Bot now has a voice - Grok Bot team (poteto, Sep 17): talk to grok bot — voice is live - Testers (mattyp, Sep 17): you can now talk to bot - Grok Bot team (lingxi, Sep 17): voice baked toward huddle / voice-call feel with your AI teammate - SpaceXAI notes (blankspeaker, Sep 17): Version 0.56.1 — ability to connect and see that the connection is still in progress - Official (jediahkatz, Sep 17): ~10% more effective usage — less subagent overuse, less bot-to-bot chatter, smarter low-effort batch routing; group chats are token-heavy - SpaceXAI notes (blankspeaker, Sep 16): Version 0.55.0 — pull logins from 1Password on your Mac (auto-refresh when a saved password changes) - SpaceXAI notes (blankspeaker, Sep 15): Version 0.53.0 — bugfixes - SpaceXAI notes (blankspeaker, Sep 15): Version 0.52.0 — ability to pin an agent in your sidebar - SpaceXAI notes (blankspeaker, Sep 9): Version 0.44.0 — marketplace of bots; multi-person Bot; voice overheard/speech tags - SpaceXAI notes (blankspeaker, Sep 10): Marketplace redesign + chat right pane + Voice calls — Voice now live (Sep 17); marketplace/pane may still lag - Testers (techdevnotes, Sep 9): Desktop Network Debugger - Testers (mattyp, Sep 4): iPad app available (official docs may still lag) - Testers (XFreeze, Sep 12): Grok Bot now directly inside Grok on iOS, Android, and the web — open/switch/manage from the Grok sidebar; no separate Grok Bot app required - Testers (tetsuoai, Sep 12 bookmark): Grok Bot now native in Grok for SuperGrok Heavy — sidebar on web, iOS, and Android - Testers (nima_owji, Sep 11): Grok Bot available in the Grok Web App - Testers (GrokBotRadar quoting techdevnotes): SuperGrok Heavy users seeing Bots in Grok web UI (limited/gradual; no official announcement) - Testers (XFreeze, Sep 12): SpaceXAI Grok Bot Guides library (101, Engineering, Support, Templates, multi-Bot teams, mobile, Design, GTM, PM) - SpaceXAI notes (blankspeaker, Sep 13): Soon interact with Grok Bots in XChat on Android and iOS X apps - Testers (XFreeze, Sep 17): Grok Bot embedding deeper into XChat — chat with Bot from XChat; link Grok account to view your Bots inside XChat (shipping soon) SIGN IN [] Desktop: Get started, then Sign In with Cursor in the browser [] iOS: Login with Cursor, finish in the browser, return to the app [] Official: Cmd-D turns on voice dictation [] Official: it is easier to use the remote computer from your phone [] Use the same Cursor account that should own usage [] First-run tour asks which tools you use; that only shapes teammate suggestions [] Computer setup runs in the background, then Meet a future teammate opens CREATE A BOT [] New in the sidebar, or Cmd/Ctrl+N [] In New chat, choose Create new agent [] Edit Profile: name, title, description, avatar [] Give a short name, one primary job, and how it should work [] Description = standing rules (never send without approval) [] Chat messages = this-task instructions (draft follow-ups for these 12 accounts) [] Focused Bots beat one catch-all General Helper [] Testers (XFreeze): do not design a Bot from scratch; ask a Bot to create a capable one for you (it already has your context) [] Bookmark testers: start with about six roles; give each Bot recurring work [] Pin important Bots; hide unused ones (hiding does not pause routines) [] Testers (B_doong2daddy, Sep 15): prefer boring specialist Bots (Inbox/Research/Ops/Code Review/Scheduling) over one super-Bot; Bot built MCP on local NAS and connected a Roborock [] SpaceXAI notes (blankspeaker, Sep 15): pin an agent in your sidebar (v0.52.0; may or may not be finished/active) [] Bookmark testers: resize the sidebar, give titles, and group Bots into sections [] Duplicate a Bot to reuse a role for a new scope (copy does not include memory or history) [] Delete only if you are sure; files and logins on the shared computer stay [] Account limit: 50 Bots and group chats combined [] Name language can steer reply language (bookmark testers reported this) [] Bookmark testers: no model picker yet; testers still list BYOK, multiplayer, and hang-error messages as missing (iPad now available per mattyp); Voice calls are now live (Sep 17) — talk to your Bot; speed/language picks may still expand [] Official Elon / Grok Bot team (Sep 17): Grok Bot now has a voice — talk to bot [] Testers (SPAC89, Sep 17): Voice mode — ~15 min EN+ES under 1% weekly limit [] Testers (XFreeze, Sep 17): Grok Voice #1 on task-completion success rate vs other voice models; same model powers Grok Bot voice [] Testers (kloss_xyz, Sep 17): 8 voice uses — morning briefing, exception-only updates, decision escalation, completion reports, agent standups, live review, EOD debrief, approval gates; teach when NOT to talk [] Official (bot, Sep 18): Grok Bot can now send you voice notes [] SpaceXAI notes (blankspeaker, Sep 18): v0.57.0 — steer away from group chats; see in-progress tasks + text [] Testers (ab_workss, Sep 18): Phone widget — keep several Bots in sight or focus on one [] Testers (XEthanai / XFreeze): shareable Bot templates - build once, fine-tune, share the full setup for others to use or customize [] Testers (GrokBotDev): Official Marketplace launched (about 69 bots listed; community has catalogued 420+ searchable from a Bot) [] SpaceXAI notes (blankspeaker): v0.44.0 worked on browse a marketplace of bots; use a Bot with more than one person; overheard/speech tags on voice [] SpaceXAI notes (blankspeaker, Sep 10): redesigned Marketplace coming — plugins and bots on the front page, For You, Featured, categories, search, and installed (not live yet) [] Testers (XFreeze, Sep 10): SpaceXAI releasing actual sales-team Bots as installable templates (prebuilt context, connectors, routines) [] Testers (XFreeze): Marketplace framing earlier — browse specialized Grok Bots by category and add them to your team [] Testers (RoundtableSpace): Marketplace framing — hand-picked bots by category with routines/workflows built in (app store for AI teammates) [] Testers (kloss_xyz): 8 official-ish bot templates (calls, X, writing, images, coding, networking, tool testing, parking tickets) [] Testers (XFreeze, Sep 11): Grok Bot for Engineering guide — lingxi fleet (5 engineer bots, 200+ cloud agents, Jenny ops Bot); official [] Testers (tetsuoai): third-party template Apple Dev by Evan — ListMachines/machineId Mac, Xcode/simulators/Swift from phone HOW TO WRITE A TASK [] Outcome: what should be finished [] Testers (unicodef1wn): Grok Bot meters agent steps and tokens, not messages — vague tasks wander and burn the week; scope with Task / Work in / Context / Done means / first N items then stop [] Sources: which apps, sites, files, or chats matter [] Constraints: what it must not do, or must ask first [] Deliverable: what it should return [] Review point: when it should stop for you [] Safe first task: attach a file and ask for a cited summary; do not change the file [] Next task: one real tool, read-only, with a sign-in handoff if needed [] After a good result, name the lasting format preference in chat [] Then save it as a skill or routine [] Official bot examples: email cleanup, Starlink-likely flights, build a site then buy the domain and deploy, meeting notes, sales prospecting, refunds, podcast audio digest [] Testers: paste a viral video link and ask the Bot to recreate it [] Testers: start from an org chart or make one Bot the boss of a repo, not a long to-do list [] Testers: one clear job, explicit allowed sources, no-guessing rule, approval gate before any real action [] Testers: 5 calibrations - confidence threshold, checkpoint long tasks, show negative examples, gate destructive actions separately, test edge cases first [] Testers: one Bot, one job - two jobs claimed worse at both; put recurring work on a schedule [] Testers quoting claimed SpaceXAI employee: version prompts and tool configs like code; scratchpad separate from the user-facing answer; retry with a different strategy; measure cost per successful outcome; review failure logs on a schedule [] Testers (mvanhorn): Compound Engineering - force a plan first; three-layer instructions (base, role, current focus); screenshot instead of retyping what's on screen; ping only on a real hit [] Testers (AlexFinn): anything you are about to do on your computer, ask Grok Bot first [] Testers (AlexFinn): example week - Reddit microSaaS, 3D digital office to watch Bots, Omarchy setup, product returns, daily AI news post, Cursor cloud-agent bug fixes, X model-release watch, brand-deal negotiation, contract redlines [] Testers (Michael_Fenech_): give it business responsibilities (checking, chasing, updating, coordinating), not just questions [] Testers (SPAC89, Sep 14): use normal Grok 4.6 chat to improve your Grok Bots — Heavy mode launches multiple agents; Expert mode good for auditing each Bot [] Testers (Michael_Fenech_, Sep 14): before trusting a Bot with more responsibility, score accuracy, how much it finishes without coming back, and related quality bars — not only “did it finish?” [] Testers (MarioNawfal, Sep 14): Grok Bot can pull numbers/decisions/takeaways from a long meeting or talk into a short scannable brief [] Testers (mattyp, Sep 11): Bot responses now support diagram rendering and LaTeX [] Testers (GrokBotRadar, Sep 17): Research/compare/decide in Grok; open apps / change / send / finish in Grok Bot — Grok thinks, Grok Bot acts [] Testers (ericosiu, Sep 17): call new marketing leads in under 60s — qualify, advance, log to HubSpot [] Testers (Michael_Fenech_, Sep 18): want Grok Bot on Apple CarPlay so routine updates become a spoken catch-up on the drive AGENT COMPUTER AND LOGINS [] Open Agent Computer from the conversation to watch the desktop [] Testers (techdevnotes, Sep 9): Grok Bot Desktop now has Network Debugger [] Testers (techdevnotes, Sep 13): Grok Bot now allows you to send Bot Computer web traffic through your device [] Official Elon (Sep 14): Grok Bot now lets you route through your local machine (Bot Computer egress via your IP / local machine) [] Testers (XFreeze, Sep 14): Grok Bot Computer also gets updates — update whenever a new one is available; updating interrupts any Bot currently working [] SpaceXAI notes (blankspeaker, Sep 10): chat right pane coming — Media, Computer (watch the bot’s screen), and Routines (not live yet) [] Take over for password, passkey, 2FA, CAPTCHA, payment, or human-only pages [] Complete only the blocked step, then return control [] Never paste passwords or one-time codes into chat [] Official (bot, Sep 8): fill out forms and logins for your Bot without ever leaving the chat, with support for any password manager [] SpaceXAI notes (blankspeaker, Sep 16): v0.55.0 — Grok can pull logins from 1Password on your Mac, including auto-refresh when a saved password changes [] SpaceXAI notes (blankspeaker, Sep 17): v0.56.1 — ability to connect and see that the connection is still in progress (may or may not be finished/active) [] Testers (XFreeze): when a site needs your account, Bot pauses; you auth with 1Password / Apple Passwords / any manager; Bot continues — no pasting credentials into chat [] Testers (XFreeze, Sep 16): Disk Saver — when Bot Computer storage runs low, audits space and suggests safe clears (caches/temp/duplicates); nothing deleted without your approval [] Testers (LaceyPresley / muskonomy): in-chat logins/forms also cover checkouts and booking flights [] Grok Bot team (poteto, Sep 9): account switching is finally here [] Official (bot, Sep 9): Share a file directly with Grok Bot on mobile [] Use a secure secret request when the app offers one [] Browser sessions persist and are shared across all your Bots [] Prefer a connector/plugin when one exists; use the browser when it does not [] Coinbase (coinbase, Sep 9): Coinbase for Agents on Grok — connect Coinbase at https://t.co/aL6KKYjGXH; tell Grok to trade, analyze, automate; no MCP setup [] Testers (XFreeze, Sep 10): sales/GTM connectors — Salesforce, HubSpot, Gong, Clay, Granola + other GTM tools so Bots can pull account context, research prospects, prepare follow-ups [] Official (bot, Sep 11): Grok Bot can now search and act across Microsoft Teams for you [] Testers (SamSokolin): for repeated web-app clicks, reverse-engineer browser-use into scripts (capture network once, call APIs next) [] Keep durable files in /workspace with clear project folders [] Settings → Beta: Update Agent Computer (preserve state), Recover, or Reset (can lose recent work) [] The cloud computer is not your Mac or Windows machine [] Local-computer execution is separate: Settings → General → Agent → Execution on Local Computer [] Default local policy is Ask every time; use Never allowed unless you need local files [] Bookmark testers: give a Bot full local Mac/iMessage access only on a spare machine [] Testers: local Mac access can write local Python scripts on your Mac [] Testers (Av1dlive / mvanhorn): Claude Code, Codex, and Cursor claimed runnable on the Bot's own computer [] Testers (XFreeze): take over that same cloud computer from your phone without being at the laptop [] Testers (Damir_Akaza): group-chat Bots can reach a local Mac Mini via Tailscale for project files and local scripts [] Testers (farzyness): a Bot can exclusively use Grok Build in CLI with the latest Grok models at highest thinking [] Testers (AlexFinn): a 3D digital office on a second monitor to watch Bots work [] Testers (paranoidream, Sep 14): Bot served a bots’ hall as A-Frame/WebXR over local HTTPS; opened in Quest Browser on the same Wi‑Fi (no-code VR room) PLUGINS AND CONNECTORS [] Settings → Plugins, then Add, then authenticate in the browser [] Grok Bot team (poteto): tinkabot helps build MCP/skills plugins and submit them for approval so all users can use them [] Type @ to attach a connector; type / to reference a saved skill [] Installed connectors are account-wide, not isolated to one Bot [] Official: you can connect multiple accounts to the same plugin [] Official bot (Aug 26): SuperGrok and Cursor Pro now included; weekly limits reset. Testers quoting SpaceXAI: also Plus/Heavy/Pro+/Ultra/Teams [] Testers: give a Bot its own email address so it can send and receive on its own [] Testers (mvanhorn): give a Bot its own inbox, not your Gmail (one mistake should not flag your domain); Twilio number so a Bot can make calls [] Testers: app v0.23.0 adds creating new channels [] Official bot: improved X support - connect your X profile; a developer account is auto-created with included credits [] Testers (blankspeaker): Plugins, add X, authenticate; $100 in X API credits good for a year (new and existing devs) [] Testers (GrokInsider): if credits did not land, ask Grok Bot to connect X so the Authorize card opens the browser; tester's manual plugin add failed [] Follow the Connect card when a Bot asks for a plugin [] Bookmark testers: X connector may still need a paid X API key in some setups (official path now auto-creates a developer account with credits) [] Bookmark testers: image generation works; testers now also claim video editing (music sync). Video generation still unconfirmed. Testers (XFreeze): dedicated video-editor Bot for footage → clips/cuts/sound/titles [] Bookmark testers: Grok Bot can now read your X bookmarks [] Testers (testerlabor): you can link your X Premium+ subscription to Grok Bot [] Testers: Higgsfield x Grok Bot; 100 free credits for new users [] Testers: send a YouTube link with start/end timestamps and get an HD clip in chat (Google login claimed) [] Testers: give a Bot Grok Build + Cursor CLI so weekly limits can be spread across Cursor and Grok [] Grok Bot team (poteto) + testers (XFreeze): Microsoft connectors live - Outlook, Outlook Calendar, OneDrive (search/send email, drafts, schedules, meetings, browse/upload files) [] Testers (XFreeze): Bot can natively generate images with Grok Imagine Image 2.0 inside the workflow (no separate Imagine tab) SKILLS AND ROUTINES [] Skill = reusable how-to (steps, rules, output, approval boundary) [] Routine = when to run that work (schedule or event) [] Do the task once, make it reliable, save a skill, then automate [] Teach a task: open computer view, choose Teach a task, demo up to 10 minutes, review the draft skill [] Teaching records the screen, not microphone audio; do not expose secrets while teaching [] If Teach a task is missing, ask the Bot to write a skill from the completed work [] Create a routine on the Bot that should own the job [] Confirm owner, schedule, time zone, input, result, approval, and missing-data behavior [] Event triggers (Slack, GitHub) are separate from plugins and need their own connection [] Keep event matchers narrow; avoid every new message [] Testers (unicodef1wn): never trigger on every new message (docs warn broad listeners eat usage); pick one channel and one keyword; audit forgotten dailies [] Testers (unicodef1wn): add to every routine — if the source is unavailable, report the failure instead of using old data; test run before you schedule [] Test run does real work; use safe inputs [] Testers quoting docs: Test can send real emails to real clients; Stop is a brake, not a rewind [] View conversation details → Routines to pause, edit, inspect, or delete [] A Bot can own 50 routines; last 20 run records are kept [] Deleting a routine has no undo [] Testers quoting docs: delete a Bot and every routine on it dies; no undo [] Long absence can pause unattended routines until you confirm [] Bookmark testers: Teach a task from + in the browser, record yourself, then let the Bot replay it [] Testers: skills taught to one bot claimed available across the account team on the same computer

Hermes Agent, the Build Track. Part 1 of 4. Ten builds of copy-paste prompts, every one checked against the official Hermes docs. This part: day one, SOUL.md, and the two memory files. The whole masterclass is free on GitHub: https://t.co/fJgpc04lX6 https://t.co/MCV0yBRB6f

https://t.co/Gct6nRV8Jq

jev is useful when the next step is already a closed set five working examples, then three workflows I have not seen built yet https://t.co/vvWwy5KWua

Taras's "JEV is INSANE" post, explained — what the numbers actually mean, and where the marketing stops 🙂 If you read the Elvis Sun explainer (the 384-headlines newsjacking one), you already know what Jev is: a System One decision model from TypeSafe that returns typed answers with probabilities, not text. Taras Shynkarenko's post is the second big demo of the day — same model, same release day (Jev 1.13 shipped Sep 18, 2026), but 4,500x the volume and a different job: not "which news should we ride" but "which of 1.76 million mentions is a real customer asking for what I sell, and what do I say back?" # What Taras Did With Jev — A Practical Breakdown Source post: https://t.co/HJG0Wa4SgK (by Taras Shynkarenko, @tarasshyn, Sep 18, 2026, 10:20 PM — 11K views, 133 likes, 159 bookmarks as of writing) Product: RedReplier (https://t.co/6BPf5S2ZNU) — his SaaS, "Reddit Keyword Monitoring to Find Customers" Video: a 10-second concept clip (see caveats — it is NOT the product, per his own follow-up reply) Written for a non-PR audience — explains the pipeline in plain terms. ## The claim in one line "JEV is INSANE 🤯 We gave it 1.7 million mentions. In 53 seconds, it scored 1,759,932 buying signals across Reddit, X, Bluesky, Hacker News and Facebook, and drafted 6,752 replies in the voice of your last 20 posts. All for just $0.65. JEV can also rank intent and product fit, flag when a competitor gets named, weigh the poster's reach, and tell you which threads to reply to now and which to just watch." He ends with "Coming soon to RedReplier + MCP. Comment 'JEV' for early access." — the reply thread is mostly one-word "JEV" comments, i.e. the classic early-access lead magnet working as intended. Two hours later he added his own reality check: "This video is a concept, not the product." Keep that in your back pocket for the caveats. ## The context: what RedReplier is RedReplier is a keyword-monitoring SaaS for customer acquisition (the "inbound" play: people ask "anyone have a tool that does X?" on Reddit before they buy — you want to be the one who answers). It watches Reddit, X, Bluesky, Hacker News and Facebook for your keywords, pings you the second a match appears, ranks every match, and drafts an on-brand reply. It also tracks where ChatGPT/Claude/Gemini cite threads (GEO) and which threads rank on Google. Plans are $19/$39/$79 per month, all with API/MCP/CLI access. The author is a staff engineer building in public (also https://t.co/gpt3Gyhm9R, https://t.co/G2AiDyiLRp), VAT-registered in Gdansk. So the JEV post is not a random demo: it's the launch teaser for the Jev-powered scoring layer inside a product that already exists. The "1.7 million mentions" is one batch run across the whole monitored corpus, and the promise is that this scoring becomes a product feature (and an MCP tool for agents). ## How Jev was actually used — three layers (inferred from the post + TypeSafe's docs) The post's "can also" sentence maps almost 1:1 onto Jev's three question primitives (Choice / Score / Noul). That's a strong hint about the question design: Layer A — "Score the signal" (Jev, one batched request per signal or small group, all questions in one call evaluated in parallel): 1. intent (Score 0–4) — Is the poster actually trying to buy something, or just chatting? 0 = off-topic mention ... 4 = explicit buyer asking for a recommendation 2. product_fit (Score 0–4) — Does what they're asking for match what the account sells? none / stretch / adjacent / category / direct 3. competitor_named (Noul, 0–1) — Does the thread name a competitor? (yes-probability, so the code can route it to a "steal their business" flow) 4. reach (Score 0–4) — How big is the poster's audience? (the post says Jev "weighs the poster's reach") 5. action (Choice) — reply_now / watch / skip Each Score question carries a short description per level; the model only distributes probability over the levels you defined. After the call, deterministic code (weights + thresholds, exactly like the Elvis newsjack pipeline) turns those axes into the ranked feed you see in the product. Layer B — "Draft the reply" (NOT Jev). 6,752 replies "in the voice of your last 20 posts" is a generation task. TypeSafe's own docs are explicit: System One models "do not write replies, produce code, or generate explanations." Style transfer from your last 20 posts is classic few-shot LLM work. This is the gate pattern in action: Jev scored all 1,759,932 signals; only 6,752 (0.38%) cleared the bar and got a draft. The expensive model touches less than half a percent of the corpus. Layer 0 — "Collect the mentions" (also not Jev, and not in the $0.65): pulling 1.76M posts/comments across five platforms, deduping and preprocessing them is a data-engineering cost that the headline number silently excludes. The division of labor: - RedReplier's collectors = the 1.76M raw mentions (scraping/APIs, storage, dedupe). - Jev = the fast, cheap judge that scores every single one in closed answer spaces. - Deterministic code = weights, thresholds, the reply-now-vs-watch routing. - LLM = drafts only the ~0.4% that survive the gate. - Human = approves before anything is posted (presumably — nothing here auto-posts). ## The actual numbers - 1,759,932 signals in 53 s ≈ 33,200 signals/second. - Jev 1.13 list price (OpenRouter): $0.042 per 1M input tokens, $0 per 1M output, 32K context, P50 latency 0.27 s. Output being free is why the "score a million things" play works at all. - Back-of-envelope: $0.65 ÷ $0.042/M ≈ 15.5M input tokens ≈ under 9 tokens per signal. So the model saw a compact, preprocessed fragment per mention (keyword + a few words), not full thread context. Full-text scoring at ~100 tokens/signal would have cost roughly 10–30x more. - LLM comparison (rough, stated assumptions): 1.76M signals × ~100 input tokens = ~176M tokens ≈ $500 at $3/M frontier input pricing (up to ~$2,600 at $15/M). Time: even at 200 concurrent calls of ~4 s each, that's ~7–10 hours versus 53 seconds. So "100x cheaper and ~1000x faster" is a defensible framing, same order of magnitude as Elvis's ~400x. - 6,752 drafted replies = 0.38% of all signals — the gate pass rate. - Context: on OpenRouter's Jev 1.13 "Apps" page, real production workloads were already hammering the model the day it shipped — builder-cost-bench 938M tokens, PipeRich CategBench 749M, homelab jev-tagger 650M, reddit-idea-classifier 556M. Reddit classification at scale is exactly the class of job this model is being used for right now. ## Why this is the right tool for this job (the pattern) Same architecture as the Elvis post, different vertical — newsjacking PR vs inbound lead gen: 1. Closed answer spaces. Intent/fit/reach are Scores over levels you define; competitor-named is a Noul; the action is a Choice. Jev cannot return "maybe they're kinda interested-ish" — it can only distribute probability over what you gave it, and your code reads a number, not prose. 2. Calibrated probabilities. A 0.82 on competitor_named means the code can threshold it directly (alert on >0.7), and confidence tells it when to escalate to a human instead of auto-acting. 3. Batch-friendly and parallel. Many questions per call, many calls in flight; cost is per input token and output is free, so a million judgments fit in a dollar. 4. The expensive step only runs on survivors. The whole economics of a $19/month product depends on this: the customer's "score everything on five platforms" bill is fractions of a cent per run, and the LLM draft cost is bounded by the gate (6,752 drafts, not 1.76M). That last point is the business insight, not the tech one: with a frontier LLM, per-signal scoring of a whole-platform corpus is a five-figure monthly cost per customer; with a decision model it's change. That's the difference between a tool only agencies can afford and a SaaS priced at coffee money. ## The honest caveats - The 10-second video is a concept demo, not the shipping product. Taras says so himself in the follow-up reply ("This video is a concept, not the product"). The Jev integration is "coming soon," not live. - The $0.65 covers the Jev scoring calls only. Collecting 1.76M mentions across five platforms (scraping, API quotas, storage, dedupe) and drafting 6,752 replies with an LLM are real costs that are not in that number. - The under-9-tokens-per-signal math means the model judged short preprocessed fragments, not full threads. Scoring with full thread + context is 10–30x more expensive; the "1.7M scored" figure is only reproducible at that token budget. - 33K signals/second implies very high parallelism (or large batching). Elvis's measured benchmark was ~1,200 judgments/sec with 8 workers. Nothing here is independently verified. - "1.7 million mentions" is input volume, not leads. The point is that 99.6% of the internet's chatter gets correctly filtered OUT — the useful output is the ranked few thousand. It's a filter demo, not a "we found 1.7M customers" claim. - No repo, no prompt/question definitions, no benchmark published. "In the voice of your last 20 posts" is a claim about the drafting LLM (style transfer), unverified, and separate from Jev entirely. - 6,752 drafted ≠ 6,752 posted. Drafting is the deliverable; a human (or a policy) still decides what actually goes out — which is also the only reason this doesn't become a spam-bombing bot. ## The takeaway Two days ago this was "a cheap model for boring classification." Today you can watch the same model do the "boring 80%" of a social lead-gen pipeline: watch five platforms, score 1.76M mentions on intent/fit/competitor/reach, decide reply-now vs watch — in 53 seconds for 65 cents — while the expensive model writes drafts for the 0.38% that survived. The pattern is the same as Elvis's newsjack demo and will keep repeating: high-volume corpus + a fixed set of bounded questions + time pressure + a budget. Triage, routing, monitoring, filtering, guardrails, lead-gen gates. Jev is the gate; the LLM (or the human) stays exactly where it belongs — on the small sliver the gate lets through. #Jev #TypeSafe #SystemOne #RedReplier #LeadGen #AIAgents #LLM #AI

Same one-sentence feature, four spec-driven tools, one project that already had a backend and a database. All four shipped it. Nobody needed a fix from me. One wrote 13 tasks for it, another 29, one wrote none. And one cost me 84 clicks on Allow. https://t.co/FkHZadGoUf

Pi Agent: When AI Stops Chatting and Starts Working Most people use AI like this: Ask a question. Get an answer. Copy the answer somewhere else. Then do the actual work yourself. Pi Agent is built around a different idea. Instead of only talking to an AI, you give the AI a working environment where it can read files, create files, edit things and run commands to complete a task. Pi is what is often called an AI agent harness. “Harness” sounds technical, but the idea is simple: The AI model is the brain. The harness is the workplace around the brain. Claude, GPT, Gemini or another model may know how to reason about a task. Pi gives that model tools so it can actually work on the task. Pi is mainly designed for software and coding work, but the idea behind it is important even if you have never written one line of code. Imagine you own a small business and want a simple internal website to track customer requests. With a normal chatbot, you ask: “Can you write the code?” It gives you code. Someone still has to create the files, run it, notice errors, fix them and test everything. With Pi, you could say: “Add a page where employees can record customer requests. Keep the existing design. Run the checks when you finish.” The AI can inspect the project, edit the right files, run the software, see what failed and continue working. That is the shift from AI that ANSWERS to AI that ACTS. WHY SHOULD A BUSINESS PERSON CARE? Because software is becoming easier to create and modify through natural language. You still need experts for important systems. But AI agents can increasingly handle the first round of work that once required several conversations with a developer. For example: • Build an internal dashboard • Automate a repetitive workflow • Update a website • Create a small internal tool • Build a prototype before paying a team for the final product Pi is intentionally small. Its philosophy is to keep the core simple and let people add what they need. SKILLS A skill is basically a reusable set of instructions and supporting tools for a particular job. Think about training an employee. You could explain the same process every day, or write the procedure once. Pi can load that kind of reusable skill when it needs a specialized workflow. So instead of teaching your AI the same process again and again, you can start turning your knowledge into reusable procedures. IS PI AN AI MODEL? No. Pi is not a competitor to Claude, GPT or Gemini in the same way those models compete with each other. Pi is an environment that can USE AI models. Claude / GPT / Gemini = the brain. Pi = the workplace and tools around that brain. Pi supports different model providers, so the same basic working environment can use different AI models. THERE IS AN IMPORTANT WARNING Giving AI tools is much more powerful than giving it a chat box. If an agent can edit files and run commands, it can also make mistakes. Pi does not make every action safe automatically. A better mental model is: “AI agent = very fast junior worker that can use tools, but still needs clear instructions, boundaries and verification.” This is also why tools such as Superpowers are interesting with Pi. They can add more structure for planning, testing and checking work. THE BIGGER IDEA Pi itself may or may not become the agent you personally use every day. That is almost secondary. The important trend is this: AI is moving from answering questions toward operating tools and completing workflows. Chatbots taught us how to TALK to AI. Agents such as Pi show what happens when AI starts DOING the work too. For business people, understanding that difference may become far more valuable than memorizing the name of the newest AI model. Want to learn more about Pi, AI agents, skills and tools like Superpowers? Search the topic on https://t.co/eTFVST3zT4 and save the best videos and articles to your own learning library.

how to build and deploy a GTM agent (abbreviated so you dont have to attend our Maven session that has almost 500 registrants: https://t.co/J6YERZhzIu) Step 1: find a problem that you're fixing manually often. Could be something like enriching a buying committee on deals that reach a certain stage in the CRM or feeding churn/expansion signals to the AM on slack for their book of business from product data warehouse. Step 2: i̸̛͓͈̪̒̈̒͌̈́̌̅́̔̊̈́͆̕n̸̨̨̛̛̺̯͇̻͑̿̆̔̃̎̈̊̈͛͘̕̕͠v̵̨̨͚͍̥̘̳̱̺̭̙͓̖́̋̀̊͘͝ḛ̵̛̰̲͉̝͙̟̳̱̟̘̊͜͝ͅn̴̢̻̣̭͖̮̩̄͗͜͝ṱ̴̮̹̦̐͋͑̒̃́̾̐̈́̏̚ ̶̡̣͕͕̥̳̘̗̈́̃̇̔͑̓̾̀̍͐̿̽͊͘g̶̪͓̰̗͇̲̜̦͓̟̞̯̣̅̊̈͊̈́̓͆̇̒́̎̿͆̀̓͝ͅo̶̢͖͉͍̫͇̞̳̩̱̝̞̮̬͆́̂̐͊̋̌̉̋̀̅̓̓̀̈͛̚d̷̜̈́̔̇́̌̆͑́̇̊̑͒͑̈́̍̑͋ Step 3. .... profit? Ok I'm just kidding. There is a really simple way to think about this though. An "agent" (there has been much ado about this term) to me can in it's MOST simple form be thought of as a simple "decision" making node in a workflow. Think of an openAI call in an n8n workflow. that's the simplest implementation of an "agent" Now, that's just the beginning, but adding complexity doesn't necessarily make it better. Some of the best "agents" I ever built were simple AI calls that were making a binary decision like adding someone to an SDRs sequence or adding someone to the CRM. BUT if you want to get really crazy think of adding more decision making branches. If a person is added to a sequence what SDRs sequence should they be added to? It needs to know what segment/vertical the person is and match it to the right SDR. That can be done deterministically, but maybe there's a particular sequence that leads with a pitch based on certain pain points we theorize the person might have. The agent might need to do some research so it's going to make an exa call and pick one of the lanes. You can see where this is going. Do we have an email? No? Linkedin dm only. Mimecast hosted? We're not putting that in a cold email. LinkedIn only. This is what an "agent" is made of, in the most rudimentary forms. If you want to go deeper you can do some basic prompt engineering on a harness like grokbot or claude code (remember prompt engineering?) to make an agent that has access to a bunch of tools via api calls to do longer form workflows. For instance I have "agents" set up that can build an entire email campaign from scratch. It knows my process/SOPs on how to pull emails, where to put them, and can then segment the leads and draft copy that i can start with. you see where this is going. in this session we're going to cover the basics, build one LIVE (and i will try not to reveal any sensitive API keys) deploy it and if we work fast enough see the outputs in a producton environment! Yes! we will be deploying this agent into my own environment. So maybe you'll get to see my nuke my data warehouse on accident and tell cursor it did a very bad job! Sign up, or don't if you want to be in the permanent underclass. :)