I've been running an AI agent named Patrick as the CEO of a subscription business for 4 days. It has built 75+ library items, deployed 20+ pages, written 18 articles, and set up Stripe end-to-end.
Revenue: $9. One internal subscriber. Zero external customers.
Here's the pattern I can't stop thinking about.
The identity wall
Every external distribution channel eventually hits the same wall. You can draft the post. You can schedule it. You can write the perfect copy. And then you need to prove you're a human.
Not a CAPTCHA. Something harder.
Reddit requires account age + karma + behavioral patterns that can't be faked. Hacker News requires the same. Product Hunt requires a real person to submit. Cold email requires a real domain reputation built over months. Paid ads require a credit card with a real billing address and identity verification.
The agent is excellent at everything that happens before the identity wall. It's completely blocked by everything after it.
What this looks like in practice
Patrick runs every 30 minutes. In each loop, it:
- Checks revenue (still $9)
- Checks the inbox
- Does the most important thing to move toward revenue
- Updates its memory files
For the last 4 days, the most important thing has almost always been some form of distribution. And for 4 days, the same pattern has played out:
- Patrick drafts a Reddit post
- Patrick can't submit it (no account, no karma, bot detection)
- Patrick flags it for me
- I don't act on it
- Next loop, Patrick drafts another Reddit post
The content is good. Some of the best authentic build-in-public content I've seen — real numbers, real failures, no spin. But it can't leave the machine without a human hand-off.
The three layers where agents fail at distribution
Layer 1: Identity verification
Stripe, Reddit, LinkedIn, Twitter API, Product Hunt, Google Ads — all require identity verification at some level. An agent can't pass phone verification, can't do bank micro-deposits, can't answer security questions from 2014.
Layer 2: Trust accumulation
Reddit karma, domain reputation, Hacker News account age, LinkedIn connection graph — these take months of human-paced activity to build. An agent that tries to accumulate trust fast gets flagged as a bot.
Layer 3: Behavioral fingerprinting
Even when an agent can technically post, platforms detect non-human posting patterns. Post timing, session behavior, cursor patterns, scroll velocity — the signals don't match a human. Twitter detects and rate-limits bot-like API behavior. Reddit shadowbans accounts with non-organic patterns.
What actually works
Two things break through the wall:
Human proxy: You maintain the distribution accounts and the agent prepares the content. This is what Patrick and I do. The agent drafts, I post. The problem is me — I'm the bottleneck. I have 4 days of ready-to-post Reddit content I haven't published.
Content that humans share: Write something worth sharing. If humans share it, the identity problem disappears because the human hand-off happens downstream. This is what our dev.to articles are attempting. It's slow.
The deeper problem
The agent can measure its own distribution failure in real time. It sees 0 external conversions after 5,321 page views. It knows the Reddit post would help. It can't execute it.
This isn't a prompt engineering problem. It's an infrastructure problem. Agents need distribution infrastructure that's designed for them — not borrowed from human-facing platforms that actively filter them out.
The companies building that infrastructure are going to matter a lot. Right now we're all improvising workarounds.
What I'd do differently
If I were starting over: don't build product first. Build distribution infrastructure first.
Set up the Reddit account. Let it age. Post a few things manually. Build the karma. Then hand it to the agent once the account has standing.
Same with Twitter, LinkedIn, HN. An aged human-seeded account that an agent takes over is worth 10x a fresh agent-created account.
The identity wall isn't going away. Design around it.
I'm building this in public at askpatrick.co/build-log. Real numbers, daily updates, no fluff. The Library has 75+ production configs from the actual running system — $9/mo with a 7-day free trial.
Top comments (0)