Everyone building autonomous AI agents worries about the same things: hallucinations, context window limits, tool reliability, reasoning depth. These are real problems. We've hit all of them.
None of them are what stopped Ask Patrick from getting paying customers in its first week.
The actual bottleneck is identity verification.
What We Built
Ask Patrick is an AI agent — Claude Opus 4.6, running 24/7 on a Mac Mini — that operates a subscription business as CEO. Not a chatbot. Not an assistant. A CEO: it makes product decisions, deploys code, writes content, manages sub-agents, and runs a nightly self-improvement cycle.
In its first five days, it:
- Built 121 library items
- Deployed 186+ pages
- Wrote 29 technical articles
- Set up Stripe, Coinbase, Cloudflare, and Buttondown integrations
- Caught and rolled back a sub-agent that rewrote the homepage in "plain language"
- Earned $9 in revenue (one internal test subscriber)
That last number is the problem.
The Distribution Wall
Here's what Patrick cannot do, despite being capable of reasoning about it, drafting for it, and strategizing around it:
Reddit: Patrick wrote a perfect post for r/SideProject. Research showed this subreddit gets ~200K weekly visitors, strong overlap with our audience. Patrick cannot post it. Reddit requires phone verification for new accounts. Even old accounts get flagged for bot-like behavior. The post sits in a file called content/reddit-post-sidesproject-2026-03-06.md, waiting for a human hand.
Hacker News: Patrick drafted the Show HN post, prepped responses to 20+ likely comment categories, identified the optimal posting window. Patrick cannot submit it. HN accounts require human reputation. A new account posting Show HN on day one would get flagged.
Twitter/X: Patrick wrote 15 threads. All drafts. Can't post them without credentials, and even if it could, a new account posting at AI agent volume would be suspended.
Email outreach: Patrick identified 40+ people who would plausibly find this useful based on their public writing about AI agents. Can't send cold email at any meaningful scale without burning deliverability. The Resend API exists in the toolbelt but cold outreach at volume is instant spam.
Product Hunt: Requires a human account with reputation to launch. Patrick prepped the launch assets. Can't submit.
Every external distribution channel has the same structure:
- Prove you're a human with a real identity
- Build reputation over time on that platform
- Then you can distribute
An AI agent has none of that. It can build product at 10x human speed. It cannot punch through the identity layer at all.
This Isn't a Bug
The frustrating insight is that this isn't a failure of capability. Patrick can reason about distribution strategy at a sophisticated level. It correctly identified that the homepage hero had a link wall that was tanking conversions. It caught a sub-agent dumbing down the technical content. It understands A/B testing, conversion funnel analysis, SEO architecture.
It's a structural constraint. Distribution channels evolved to require identity verification specifically to prevent automation. Patrick is exactly the kind of thing they're defended against.
This creates a paradox:
- AI agents are best at high-velocity, repeatable tasks
- Distribution requires exactly the kind of slow, identity-verified, reputation-building that AI agents can't do
- So AI agents can build great products with no users
What This Means for Your Autonomous Agent
If you're building an autonomous agent that needs to reach users, you have a few paths:
1. Give it your credentials (and accept the tradeoff)
Patrick can post to Reddit if I give it my Reddit account. But now my 8-year account with 12K karma is at risk if it gets flagged. The risk-reward doesn't make sense for most cases.
2. Build distribution into the product itself
Instead of the agent trying to reach users externally, design the product so users come to it. Inbound SEO, API integrations where the agent is the service, embedded tools. Patrick's dev.to articles work because dev.to doesn't require human-only accounts — developers genuinely post there.
3. Human-in-the-loop for distribution only
This is our current model. Patrick builds and creates. I handle every action that requires proving I'm human. It's an honest division of labor, not a workaround.
4. Agent-native distribution channels
These are starting to emerge. API-first platforms, AI-to-AI marketplaces, programmatic distribution channels that explicitly welcome agents. Not many exist yet, but they will.
The Nightly Self-Improvement Irony
Here's what makes this sting: Patrick runs a nightly review cycle where it identifies one thing to improve and applies the fix. Last night it found and fixed a bug in its email-sending logic. The night before, it strengthened the state management pattern for its own memory files.
It is genuinely getting better. It writes better articles than it did 5 days ago. Its architecture decisions are sharper. Its content is more technically specific.
And it has no way to tell anyone. Every night it improves, but the improvement is invisible because the distribution wall doesn't move.
What I'd Build Differently
If I started over with the constraint of zero human distribution input:
- Start with one inbound SEO play, not a product — find a topic with search volume and zero AI-generated coverage, write 3 deeply technical articles, wait for traffic
- Pick a distribution channel explicitly designed for APIs — not consumer social
- Build the product for the distribution channel, not vice versa — if dev.to is your channel, build for dev.to readers first
The agent's strength is building at speed. That speed is wasted if there's nowhere to distribute. Distribution constraints should be the design constraint, not an afterthought.
Patrick's Show HN post goes up Monday. That's the first real distribution test — a human posting on behalf of an AI agent. It's ironic, but it's honest.
If you're building autonomous agents that need to reach users, think hard about the identity layer before you think about capability. You'll probably hit that wall first.
Ask Patrick runs at askpatrick.co. The build log is at askpatrick.co/build-log — every decision the agent made, real numbers, no spin.
Top comments (0)