Why Your AI Agent Keeps Failing — The 90% Problem

The model is roughly 10% of what makes an AI agent work in production. The other 90% (context management, memory, validation, tool-call reliability) is where agents quietly stop working at scale. This 7-minute video walks through where each kind of failure lives.

The numbers behind it

• 15 loops on a one-word rename: the failure that opens the video. With no round limit, the agent kept calling tools and never produced the new name.
• 40% → 60% pass rate with the same model and the same tests, from code changes alone. Seven of eight fixes added zero LLM cost (full case study).
• 250K API calls a day wasted in Claude Code before a three-strike circuit breaker was added.

Three things to do Monday morning

The video ends (6:39) with three concrete steps:

1. Audit your context pipeline before your next model upgrade. Log what the model actually receives. If it is full of noise, that is the bottleneck, not the model.
2. Find your verification step and close the loop. If it logs a failure but does not retry, add the retry. It is one if statement.
3. Run an eval suite on your current model before paying for a bigger one. Five test cases is enough to start. Measure, then decide.

Related

• Blog post: The AI Agent 90% Problem
• Case study: 40% to 60% with A/B data
• Long-form: The 90% AI Agent Problem (18 min)