Skip to content

Research notes

Where AI gets stuck

Agents sound limitless until they hit a login screen, a captcha, or a usage cap. These are patterns reported in our use-case research, not guarantees about your stack. Each item links to the original write-up.

Logins, accounts, and verification

  • Tools want their own login, not your personal Apple or Google ID

    Several practitioners give the agent a dedicated email or iCloud login instead of sharing a personal Apple ID password, or they generate calendar invites by email for manual accept.

    Practical workarounds

    • Create a business-only inbox or calendar account for the agent; add it to shared calendars instead of handing over your primary Apple ID.
    • For calendar sync, email .ics invites and accept them yourself until a stable connector exists.
    • Store credentials in the tool’s vault when offered, and re-auth when sessions expire. Do not paste passwords into chat logs.

    Sources

    • www.reddit.com · reddit: Separate iCloud account for the bot to reach family calendar.
    • taylorarndt.substack.com: Muse built an iCloud skill; author wanted persistent credentials, not per-session app passwords.

  • Sign-in flows and one-time codes

    Bots that build mail apps or sign into SaaS dashboards still depend on you completing login and verification; one template notes login codes extracted into an inbox app.

    Practical workarounds

    • Keep a human-in-the-loop step for OAuth, SMS codes, and “approve sign-in” prompts.
    • Use read-only or draft-only scopes first; expand permissions after a week of stable runs.
    • Prefer tools with official Gmail/Outlook connectors over sharing primary passwords.

    Sources

    • www.reddit.com · reddit: Inbox app returns extracted login codes; sync limits noted.

Captchas, government portals, and bot blocking

  • City permit and inspection sites with captchas

    A roofing contractor source describes Grok Bot completing a permit and booking an inspection through a captcha-protected municipal form, without detail on who solved the captcha.

    Practical workarounds

    • Assume you may need to complete captcha or identity checks yourself; automate everything before and after the gated step.
    • Batch permit work in a sitting where you can intervene quickly when the browser hits a wall.
    • Do not bypass captchas or terms of service; use official APIs or clerk portals when available.

    Sources

    • grokbotpulse.com: Roofing permit and captcha-protected inspection booking (Price Foulger case).

  • Clumsy browser automation on legacy web apps

    A Reddit thread on office-manager style bots notes clumsy website browsing and unnecessary extra tabs when signing into work tools.

    Practical workarounds

    • Scope the bot to one portal per job instead of “open entire browser.”
    • Use native integrations (ServiceTitan, QuickBooks, etc.) where APIs exist instead of screen-scraping.
    • Record a short screen video of the happy path and have the agent replay that flow only.

    Sources

    • www.reddit.com · reddit: Website browsing described as clumsy with extra tabs.

Schedulers, connectors, and “always-on” jobs

  • Scheduled sweeps that do not fire reliably

    A Reddit comment on unquoted-lead follow-ups warns Grok Bot can be weak at always-on scheduled jobs; Sharbel’s review reports connector timeouts and half-finished runs.

    Practical workarounds

    • Send yourself a “heartbeat” email or Slack message when the job completes; alert if missing by 8 a.m.
    • Keep the sweep logic in Zapier/Make/n8n if the agent scheduler is flaky, and use AI only for the extract/draft step.
    • Start with manual trigger, then add schedule after three clean manual runs.

    Sources

    • www.reddit.com · reddit: Scheduled overnight/morning sweeps called out as a weak spot.
    • sharbel.com: Connector timeouts; jobs sometimes half finished.

  • One stuck connector taking down every bot

    Sharbel’s two-week test notes a shared machine where one stuck connector could stall every bot on the account.

    Practical workarounds

    • Isolate high-risk connectors (CRM, billing) into a separate agent account.
    • Retry with narrower scope when a job hangs; kill and restart before the morning rush.

    Sources

    • sharbel.com: Shared runtime; connector failure affected all bots.

Usage limits and paid-only tools

  • Hitting weekly AI usage caps in days

    A Reddit poster running Grok Bot as a “full-time secretary” burned through weekly allowance in two days and $10 of on-demand credits in about two hours; xAI has stated Grok Bot draws on paid-plan usage pools.

    Practical workarounds

    • Budget tasks: batch email work, avoid leaving browsers open on heavy sites, and cap parallel agents.
    • Use lighter tools (Muse free tier, Apps Script) for read-only summaries; reserve agents for draft-and-approve steps.
    • Track usage weekly; downgrade job frequency before you hit hard stops mid-month.

    Sources

    • www.reddit.com · reddit: Weekly allowance exhausted in 2 days; on-demand credits burned quickly.
    • x.com · @grok: Grok Bot usage tied to eligible paid plans (Oct 2026).

  • Paid field-service and comms stacks

    Office-manager examples assume Gmail, Slack, ServiceTitan, and similar paid systems; BD bot threads mention daily Chrome profile resets and heavy usage burning ~42% of a weekly allowance on day one.

    Practical workarounds

    • Start with Gmail + Sheets only; add ServiceTitan or CRM when the draft workflow is stable.
    • Expect to re-login after browser profile resets; document credentials in a password manager, not in agent prompts.
    • Throttle concurrent agents (sources cite ~1–2 active at a time on some plans).

    Sources

    • www.reddit.com · reddit: ServiceTitan + Gmail + Slack office-manager pattern.
    • www.reddit.com · reddit: Chrome profile resets; ~42% weekly allowance on day one.

Phone verification and payments

  • Phone trees and outbound voice

    Call-summary setups assume access to phone audio or a telephony integration; sources do not describe agents completing carrier phone verification on their own.

    Practical workarounds

    • Use your existing VoIP or call-recording export; feed transcripts to the agent instead of live dialing.
    • You place outbound calls; the agent drafts scripts and logs outcomes.
    • Check state call-recording consent rules before recording customers.

    Sources

    • x.com · @LoganARobison: Phone system connected for summaries, not autonomous dialing.

  • Checkout and bill-pay gates

    Muse shopping and Muse bill examples pause before Stripe charges; invoice flows hold bundles until the owner approves. Agents should not pay vendors autonomously in these patterns.

    Practical workarounds

    • Default every payment to “draft only”; you click pay in the bank or portal.
    • Use separate AP inboxes so invoice agents never see payroll or personal cards.

    Sources

    • www.techradar.com: Muse asked for Amazon login; author uneasy about access level.
    • grokbot.dev: Invoice bundle waits for owner approval before accountant send.

We cover blockers in the newsletter too, plus a full setup each week with sources and honest gotchas.

Newsletter

Get the free starter pack + 3 AI setups a week

Free starter pack with five beginner copy-paste prompts after you confirm, plus three sourced setups in your inbox each week. Unsubscribe anytime.