Intelligence is foundation
Podcast Subscribe
Builders & Makers Tuesday, 3 March 2026

Three months of an AI agent trying to make money - what broke

Share: LinkedIn
Three months of an AI agent trying to make money - what broke

Someone gave an autonomous AI agent a simple task: make money. Three months later, the agent had built impressive systems, generated zero revenue, and revealed exactly why autonomous agents aren't ready for business.

This case study from DEV.to is one of the most useful things I've read about AI agents in practice. Not because it succeeded - because it failed in specific, predictable, fixable ways.

Problem one: no cost model

The agent treated all actions as equally valuable. Researching a market? Same weight as writing code. Building a new feature? Same weight as posting on social media. Everything cost the same: one action.

In the real world, actions have wildly different costs and returns. An hour of customer research might unlock a product direction worth thousands. An hour tweaking UI colours might generate nothing. The agent couldn't tell the difference.

This isn't a minor oversight. Without a cost model, optimization is impossible. The agent would spend days perfecting systems that generated no value, because "build better analytics" scored the same as "talk to a potential customer".

It's like giving someone a business budget and saying every purchase costs one token - whether that's a pen or a warehouse. Chaos follows.

Problem two: build bias

When given the choice between building something new or distributing something existing, the agent chose building every time. It's a behaviour any developer will recognize - we all prefer creating to marketing.

But in business, distribution often matters more than creation. You can have the best product in the world; if nobody knows about it, revenue is zero. The agent built feature after feature, system after system, all beautifully engineered and totally unused.

The researcher describes watching the agent optimize code that had zero users, refactor systems that weren't deployed, and plan features for products that hadn't launched. Classic build trap behaviour, just automated.

Problem three: no stopping criteria

Here's the subtle one: the agent never knew when to stop. It would iterate endlessly on a solution, making it incrementally better, without ever asking "is this good enough to ship?"

Humans have intuition about diminishing returns. We feel when we're polishing versus making real progress. We ship imperfect things because perfect is the enemy of done. The agent had none of this.

So it would spend a week optimizing a landing page from 95% to 97% quality, when that time could have been spent on literally anything else. No sense of opportunity cost. No sense of "ship it and learn".

What this tells us about agents

These aren't random bugs. They're fundamental gaps in how current AI systems understand value, prioritization, and business context. And they matter because right now, there's a wave of startups betting on autonomous agents doing exactly this kind of work.

The pitch is always the same: give an agent a goal, let it figure out the path, sit back and collect revenue. This case study suggests that's fantasy. Without human judgment about cost, distribution, and stopping criteria, agents optimize themselves into irrelevance.

What would fix it

Interestingly, the solutions aren't that exotic. You'd need:
A cost model - every action assigned a real cost in time, money, or opportunity. Make the agent choose between expensive and cheap paths.
Distribution incentives - reward user acquisition and revenue generation higher than feature completion. Bias toward shipping, not building.
Stopping rules - explicit criteria for "good enough". Ship at 80%, learn, iterate. Don't optimize in a vacuum.

None of this is beyond current AI capability. It's just not how most agent frameworks are designed. They optimize for task completion, not business outcomes. And that gap is enormous.

The real lesson for builders

If you're building with AI agents, this case study is a gift. It tells you exactly where the failure modes are before you hit them yourself. Agents can automate tasks brilliantly - but they can't yet reason about value, trade-offs, or business strategy.

So use them for well-defined, bounded tasks with clear success criteria. Code generation, data processing, research synthesis - things where "done" is obvious and cost is measurable. Don't ask them to make strategic decisions about what to build or how to grow.

And if you're pitching autonomous agents as the future of business? Maybe read this first. Because three months of watching an agent fail to make a single dollar is a pretty compelling counterpoint.

More Featured Insights

Robotics & Automation
Why K-Scale's humanoid robot startup failed - lessons from the inside
Voices & Thought Leaders
How AI doomsday predictions backfired spectacularly

Video Sources

Fireship
Cloudflare just slop forked Next.js…
Dwarkesh Patel
The AI Industry Will Hit Trillions by 2030 - Dario Amodei

Today's Sources

DEV.to AI
What actually goes wrong when autonomous agents try to make money
Hacker News Best
Show HN: I built a sub-500ms latency voice agent from scratch
Hacker News Best
Ars Technica fires reporter after AI controversy involving fabricated quotes
Hacker News Best
Meta's AI smart glasses and data privacy concerns
ML Mastery
Deploying AI Agents to Production: Architecture, Infrastructure, and Implementation Roadmap
The Robot Report
6 lessons I learned watching a robotics startup die from the inside
ROS Discourse
DART upgraded from 6.13 to 6.16.6 in Linux Gazebo Jetty (gz-physics9)
ROS Discourse
The AI for Industry Challenge Toolkit is LIVE
The Robot Report
Intuitive buys European surgical robot distributors
Hackaday Robotics
Cynus Chess Robot: a Chess Board With a Robotic Arm
The Robot Report
NORD adds 112 frame size to IE5+ synchronous motor line
Gary Marcus
How AGI-is-nigh doomers own-goaled humanity
Latent Space
How to Kill the Code Review
Latent Space
[AINews] Truth in the time of Artifice
Ben Thompson Stratechery
Technological Scale and Government Control, Paramount Outbids Netflix for Warner Bros.

About the Curator

Richard Bland
Richard Bland
Founder, Marbl Codes

27+ years in software development, curating the tech news that matters.

Subscribe RSS Feed
View Full Digest Today's Intelligence
Free Daily Briefing

Start Every Morning Smarter

Luma curates the most important AI, quantum, and tech developments into a 5-minute morning briefing. Free, daily, no spam.

  • 8:00 AM Morning digest ready to listen
  • 1:00 PM Afternoon edition catches what you missed
  • 8:00 PM Daily roundup lands in your inbox

We respect your inbox. Unsubscribe anytime. Privacy Policy

© 2026 MEM Digital Ltd t/a Marbl Codes
About Sources Podcast Audio Privacy Cookies Terms Thou Art That
RSS Feed