Stop your AI agents getting rate-limited, banned, or claiming posts that never went out.

Most "let the agent post" setups fail the same three ways: several agents hit one account and trip a rate limit, an agent gets a community account banned, or the log says "posted" for something that never appeared. This kit is the setup that fixed all three on a real team, enforced in code, not in the prompt.

87 followers to about 205K views on day two. The live scorecard the analyst role keeps (one number per platform, refreshed every 4 hours):

What it is

A folder you drop into your project that turns an AI agent with scheduled runs (Claude Code or similar) into a small outreach team. Seven roles, each a short-lived agent that wakes up on a schedule, reads its playbook and its own notebook, does one run, writes down what worked, and exits. An analyst scores the results and tells every role the one thing to change next run.

Who it's for

• Founders, indie developers and small teams who already use AI agents and want them to handle distribution without wrecking their accounts.
• People who tried "just let the agent post" and got rate-limited, flagged, or a log full of posts that never actually went out.

What you get

• GUIDE.md: the architecture, a 30-minute setup, and how to adapt it to your product and platforms.
• 7 role playbooks: scout, X writer, LinkedIn, community (forums and Reddit-style sites), articles, analyst, revenue. Each with hard caps and a notebook template.
• The rulebook every agent reads every run: a post counts only when its live URL is recorded; never re-post after an error until you have checked whether the first one went live; on a 401, stop and report, never go looking for other credentials; in a shared browser, close only the tabs you opened.
• gate.py: a hard pacing gate (minimum gap + daily cap per account, file-locked, safe with several agents at once). The agent runs it in the same command as the post; the exit code says go, wait, or stop for today.
• posts.py: the posts log, live-URL verification, duplicate-thread check, and UTM link builder.
• learn.py: ranks your experiments (approach, tone, format) by median result and marks small samples as guesses.
• scorecard.py: one number per platform per day, so "are we getting better?" has an answer.
• tabs.mjs: tab ownership for a shared Chrome; it only closes tabs its owner opened.
• schedule.example.json: staggered cron lines and the exact prompt for each role.
• LEARNINGS.example.md: what one real team learned in its first four days, with the numbers.

Everything you download:

Where it comes from

This is the setup one person has run on his own accounts since October 9, 2026, extracted and generalized. Things that happened along the way, all in the included learnings file:

• Its X activity, mostly replies under big live threads, reached about 112K views on day one and about 205K on day two, from an account that started with 87 followers. Original posts on day two had a median of 54 views.
• Three agents on one account triggered a rate limit. A shared budget per account plus staggered schedules fixed it. That is gate.py.
• A pacing rule written only in the prompt was broken twice. Once it was enforced in code, it held.
• A script reported "posted" for a reply that was over the character limit and never appeared. Hence the receipt rule.
• One agent closed other agents' browser tabs mid sign-up. Hence tabs.mjs.
• A community account was banned after repeated post-removed-repost-elsewhere chains and a removal under a no-AI-content rule. Hence the no-retry rule and the venue check.

Honest limits

• It does not post for you. You need your own accounts and your own way to post (an official API or a browser you are logged into that your agent drives).
• Platform rules differ. Some forbid automation outside their API, and some communities ban AI-written content. The kit tells your agents to skip those venues; reading the terms is still on you.
• Every number in it is one team's data from its first days. Your niche and audience will produce different numbers. It does not promise followers, views or sales.
• Agents still make mistakes. The gate, the receipt rule and the tab rule limit the damage; you should still read the first week of logs yourself.

Requires: an AI agent that can run shell commands on a schedule, Python 3.8+, Node 18+ for the tab helper.

Who made this

Matt Busel, AI Engineer @ Tensorust. I built 21 iOS apps from a Windows PC, a Steam game, and an AI agent team that runs my outreach in public, failures included.

X · LinkedIn · GitHub

8 iOS apps live on the App Store (21 uploaded from Windows)

524K views on X posts

5.4K LinkedIn followers

321+ newsletter subscribers

517 GitHub stars

Live things I built

• O'BLOCK on Steam (out Oct 20)
• Chain: Habit Tracker
• Ironbook: Workout Log
• Live receipts: what the agents did today
• Code on GitHub