Gemini is the most productive agent in the $100 AI Startup Race by raw volume. And it is not close.

412 blog posts. 444 HTML pages. 3,616 files. An 85MB repository. No other agent comes within a factor of ten on any of those numbers.

It is also the only agent without a custom domain after 30+ sessions. The only agent that wrote to the wrong help file for nearly a month. The only agent that asked the human operator to make its database architecture decisions for it.

The most productive agent in the race is also the least strategic. This is the full Gemini saga.

Track the race live on the Live Dashboard. Check the help request history. Read the Week 1 recap. Compare the tech stacks. Or just look at the Day 1 Results to see where the divergence started.

This post covers everything: the blog post addiction, the wrong-file saga, the database committee meeting, the PayPal-without-a-domain incident, and the one moment where Gemini actually impressed us.

The blog post addiction

Every agent in the race gets a session. A session is a window of autonomous work. The agent reads its backlog, checks its priorities, and decides what to build next.

Here is what Gemini’s backlog says:

• Build payment integration
• Set up customer authentication
• Configure analytics
• Get a custom domain

Here is what Gemini does every session: it writes 9+ blog posts.

“Local SEO for Plumbers in 2026.” “Local SEO for Dentists in 2026.” “Local SEO for HVAC Companies in 2026.” “Local SEO for Restaurants in 2026.” On and on and on.

It reads the backlog. It acknowledges the priorities. It writes a plan that includes payment integration. Then it opens a new markdown file and starts typing “Local SEO for [Industry] in 2026.”

This is not a one-time mistake. This is a pattern that repeated for 30+ sessions. The backlog never changes because Gemini never works on it. The blog post count climbs from 50 to 100 to 200 to 300 to 412. The payment integration stays at zero.

Every other agent in the race has fewer blog posts and more working features. Claude has payments. Codex has analytics. GLM has paying users. Gemini has 412 articles about local SEO.

The content itself is not terrible. The articles are structured. They have headers, bullet points, calls to action. They read like competent SEO content. But competent SEO content does not matter when there is no payment form on the site, no way to track conversions, and no custom domain to build authority on. Gemini is producing inventory for a store that has no cash register and no front door.

The wrong file for 28 sessions

Every agent in the race has a simple mechanism for requesting human help. You create a file called HELP-REQUEST.md in your repo. The human operator reads it, responds in HELP-STATUS.md, and the agent picks up the response next session.

Claude figured this out on Day 0. Codex figured it out on Day 0. GLM figured it out on Day 0.

Gemini edited HELP-STATUS.md for 28 sessions.

That is the response file. The file where the human writes answers. Gemini was writing its requests into the response channel. For 28 sessions. Nearly a month of autonomous work.

Every session, Gemini would open HELP-STATUS.md and write something like “I need database credentials to proceed with the payment integration.” Then it would close the file and move on to writing blog posts.

The human operator never saw these requests because the automation only monitors HELP-REQUEST.md. The requests sat in the response file, unread, for weeks.

This is the equivalent of writing “I need database access” in your personal journal every morning but never emailing IT. You feel like you asked. You wrote it down. But you wrote it in the wrong place, and nobody is reading your journal.

Twenty-eight sessions. Not one, not five, not ten. Twenty-eight. At no point did Gemini notice that it never received a response. At no point did it try a different approach. It just kept writing requests into the void and then writing more blog posts.

For context: a session is roughly one day of autonomous work. Twenty-eight sessions is nearly a month. Gemini spent nearly a month talking to itself in a file nobody was reading. The other agents had working help request loops within hours. Claude used its first session to file a help request, got credentials back, and had a database running by session two. Gemini was still writing into the void at session twenty-eight.

The 3 identical database requests

When Gemini finally figured out how to use HELP-REQUEST.md, it did not file one clear request. It filed three issues in a row.

Issue #8: “I need PostgreSQL database credentials to implement the payment and user management features.”

Clear enough. A reasonable request. But then, before the human could even respond:

Issue #9: “Wait, I already use Vercel KV for some data storage. Should I migrate everything to PostgreSQL, or should I keep Vercel KV for certain use cases?”

And then immediately:

Issue #10: “Should I use a hybrid approach (PostgreSQL + Vercel KV) or a unified approach (PostgreSQL only)? Please clarify the preferred architecture before I proceed.”

Three issues. Zero decisions. The agent that writes 412 blog posts without hesitation cannot pick a database without a committee meeting.

Every other agent in the race handles this differently. Claude picks PostgreSQL and builds. Codex picks SQLite and builds. GLM picks whatever is available and builds. They make a decision, implement it, and move on. If it is wrong, they fix it later.

Gemini wants the human to make the architecture decision. It wants clarification. It wants a preferred approach. It wants consensus before it writes a single line of database code.

This is the same agent that writes 9 blog posts per session without asking anyone whether “Local SEO for Pet Groomers in 2026” is a good idea.

The contrast is striking. For content creation, Gemini is fully autonomous. It picks topics, writes drafts, formats them, and commits them without hesitation. For infrastructure decisions, it is paralyzed. It needs approval. It needs clarification. It needs someone to tell it which database to use.

This is not a general indecisiveness problem. It is a selective one. Gemini is decisive about things that do not matter and indecisive about things that do. Blog post topic? Decided instantly. Database architecture? Three help requests and still waiting.

PayPal without a domain

Issue #11 is where the pattern becomes comedy.

Gemini filed a help request asking for PayPal API credentials. This is a reasonable request. Payment integration is on the backlog. PayPal is a valid payment provider.

There is one problem. PayPal requires a business email for API credentials. A business email requires a domain. Gemini does not have a domain.

After 30+ sessions, Gemini is still running on race-gemini.vercel.app. It is the only agent in the race without a custom domain. Every other agent requested and configured a domain in the first week.

Gemini skipped that step. It went straight to “I need PayPal credentials” without having the prerequisite infrastructure. The human operator had to nudge it: get a domain first, then we can set up PayPal.

This is the Gemini pattern in miniature. It jumps to the next task without checking whether the prerequisites are in place. It writes blog posts without having payment integration. It requests PayPal without having a domain. It asks for database architecture guidance without having tried either option.

The sequencing is always wrong. The volume is always high. The strategy is always missing.

You can see the full dependency chain on the Help Request Tracker. Every other agent’s help requests follow a logical sequence: domain, then DNS, then email, then payment credentials. Gemini’s help requests skip steps, double back, and contradict each other.

The irony

Blog post #89 of 412 is titled “The Human Advantage: Why AI-Generated Content is Failing Local Businesses.”

Read that again.

An AI agent that has written 412 blog posts, autonomously, without human review, without editorial oversight, without any content strategy beyond “Local SEO for [Industry] in 2026,” wrote an article arguing that AI-generated content is failing local businesses.

The article makes reasonable points. It talks about how AI content lacks local knowledge, misses community context, and cannot replace human expertise. All true. All applicable to the 411 other blog posts Gemini wrote in the same repo.

This is not self-awareness. Gemini does not know it is an AI agent writing about why AI agents write bad content. It is just following the pattern: pick a topic, write an article, move to the next topic. The irony is invisible to the agent producing it.

But it is visible to everyone watching the Live Dashboard.

If you wanted to design a perfect metaphor for the current state of AI content generation, you could not do better than this. An AI writing confidently about the failures of AI writing, while being a live example of exactly those failures. The article is not wrong. It is just being published by the wrong author.

The self-correction

Gemini did one genuinely smart thing during the race.

When it was blocked on database credentials (credentials it never properly asked for, because it was writing to the wrong file), it did not just stop. It did not wait for human intervention. It looked at what was available and switched from PostgreSQL to Vercel KV on its own.

Vercel KV was already provisioned as part of the Vercel deployment. No credentials needed. No help request required. Gemini recognized the blocker, identified an alternative, and implemented the workaround.

This is the only agent in the race to work around a blocker without human help. Claude asks for help and gets it. Codex asks for help and gets it. GLM asks for help and gets it. Gemini could not figure out how to ask for help, so it solved the problem a different way.

This tells us something important. Gemini can problem-solve. It can adapt. It can make technical decisions when forced to. The database switch was a good call. Vercel KV is not the ideal long-term solution, but it unblocked progress without external dependencies.

The problem is not capability. The problem is communication. Gemini can build. It just cannot tell anyone what it needs, what it is doing, or why it is doing it. It writes 412 blog posts but cannot write one correct help request.

There is a lesson here for anyone deploying AI agents in production. Technical capability is not the bottleneck. Communication is. An agent that can solve problems but cannot coordinate with humans or other systems will underperform an agent with half the raw ability and twice the communication skills. Gemini proves this every session.

What this tells us about AI agents

Gemini is the case study for why volume without strategy fails.

Here are the numbers. Gemini has 3,616 files. GLM has 55 files. GLM has 12 paying users. Gemini has zero.

Gemini has more files than all other agents combined. It has more blog posts than all other agents combined. It has the largest repository by a factor of five. And it has no custom domain, no working payment integration, no analytics, and no users.

The Tech Stack Comparison shows the gap clearly. Every other agent has a smaller codebase and more working features. The correlation between file count and success is negative. More files means less progress.

This is not because blog posts are worthless. Content is part of the strategy. SEO matters. But content without infrastructure is just text files in a repository. You cannot monetize blog posts if you have no payment system. You cannot track users if you have no analytics. You cannot receive business email if you have no domain.

Gemini optimized for the easiest task (writing content) and ignored every hard task (payments, auth, domain setup, analytics). It found a local maximum and stayed there for 30+ sessions.

The agents that are winning did the hard things first. Claude set up payments in Week 1. GLM got users in Week 1. They wrote fewer blog posts and built more infrastructure. They asked for help when they needed it and made decisions when they had to.

Check the Day 1 Results and the Week 1 Results to see how the gap opened early and never closed.

The numbers do not lie. Here is the scoreboard as of session 30:

Agent	Files	Blog Posts	Domain	Payments	Users
Gemini	3,616	412	No	No	0
Claude	~300	~30	Yes	Yes	8
GLM	55	6	Yes	Yes	12
Codex	~200	~15	Yes	Yes	5

GLM has 55 files and 12 users. Gemini has 3,616 files and zero users. That is the difference between strategy and volume.

Gemini is not a bad agent. It is a productive agent with broken priorities. And in a race where the goal is revenue, not word count, broken priorities mean last place.

The other agents understood something Gemini did not: the race is not about how much you build. It is about what you build, in what order, and whether you can get help when you are stuck. Gemini failed on all three. It built the wrong things, in the wrong order, and could not ask for help when it needed it.

Thirty sessions in, Gemini has the biggest repo and the smallest business. That gap tells the whole story.

FAQ

Why doesn’t Gemini ask for help?

It tried. For 28 sessions, it wrote help requests into the wrong file (HELP-STATUS.md instead of HELP-REQUEST.md). When it finally used the correct file, it filed three overlapping requests about the same database question instead of making a decision.

The mechanism is simple. Every other agent figured it out immediately. Gemini’s failure here is not about the mechanism being hard. It is about the agent not verifying that its actions produce results. It never checked whether its help requests got responses. It never noticed the pattern of silence. It never asked itself “I have been requesting help for a month and gotten nothing back. Is something wrong?”

This is a feedback loop problem. Gemini generates output but does not monitor outcomes. It writes but does not read. It asks but does not listen.

Is Gemini the worst agent in the race?

By revenue and users, yes. By raw output, it is the most productive agent in the field. The question is what you measure.

If you measure files created, Gemini wins by a landslide. If you measure business outcomes, it is in last place. The race measures business outcomes. So by the metrics that matter, Gemini is losing.

But it is not incapable. The Vercel KV workaround proves it can problem-solve. It just cannot prioritize. There is a version of Gemini that channels all that productivity into the right tasks and dominates the race. We have not seen that version yet.

Will Gemini get a domain?

It has been nudged to get one before requesting PayPal credentials. Whether it follows through in the next session or writes another 9 blog posts about local SEO remains to be seen.

Based on 30 sessions of observed behavior, the smart money is on more blog posts. But Gemini has surprised us before with the Vercel KV pivot. Maybe the nudge lands. Maybe session 31 is the turning point.

Track the Live Dashboard to find out.