Why AI Agents Hit the Wall

· #today-i-learned, #ai, #agents, #engineering · 2 min read

I found this excellent article by Utkarsh Kanwat: Betting Against Agents.

Compounding Works Both Ways #

We usually talk about compounding in positive terms—interest accumulates, wealth grows, skills build on skills. But with AI agents, compounding works against you: the failure rate of an agentic process compounds with each step.

This framing really clicked for me. I’ve noticed that I need to break problems into small steps and have agents implement incrementally, verifying each step. The article gives intuition into why: even with 95% reliability per step, you’re down to 36% success rate after just 20 steps. Not good enough for production systems.

Error compounding in multi-step AI agent workflows Error compounding visualization from the original article

That said, I’m not convinced this is unsolvable. Starting with clear specifications, building in feedback loops (like test runs), and giving agents self-verification capabilities could help. But the article makes a compelling case that fully autonomous agent approach is fundamentally flawed—especially when those verification steps themselves become prohibitively expensive.

The Triple Threat #

The article identifies three core problems:

  1. Error Compounding: Math doesn’t lie. Multi-step workflows exponentially degrade in reliability.

  2. Token Economics: Context windows create quadratic cost scaling. A 100-turn conversation can cost $50-100 in tokens alone.

Token cost scaling in AI agent conversations Token cost scaling visualization from the original article

  1. Tool Engineering Complexity: About 70% of agent work is building robust tools—handling failures, complex state, authentication flows. This isn’t an AI problem, it’s a systems engineering problem.

What Actually Works #

Instead of fully autonomous agents, the author advocates for:

  • Constrained, domain-specific tools
  • Clear boundaries and human oversight
  • Solving for reliability over autonomy
  • Building on solid software engineering foundations

Their prediction? Venture-funded “fully autonomous agent” startups will hit the economics wall in 2025, while pragmatic, bounded approaches succeed.

This is a summary of “Betting Against Agents” by Utkarsh Kanwat.