Every cohort we run, someone ships a demo on day three that genuinely impresses people. The agent works. The latency is fine. The outputs look clean.
By week four, that same system is broken in production.
This is not bad luck. It is the same failure every time.
The demo is not the system
A demo proves the happy path. You control the input, the context is fresh, the API keys are valid, and you are watching it run. Of course it works.
A production system has to survive:
- Inputs you did not anticipate
- Context windows that fill up mid-conversation
- API rate limits hit at 2am
- A model update that changed the output format
- A user who pastes in 40,000 tokens of their codebase
None of these appear in a demo. All of them appear in production.
The five failure modes we see every cohort
1. Hallucination with no guard The model returns a confident, plausible, wrong answer and the system passes it downstream as fact. Fix: output validation, structured schemas, explicit confidence thresholds.
2. Runaway spend No token ceiling, no request budget, no alerting. One bad input triggers a loop and runs up $80 overnight. Fix: hard limits at the API call level, not the application level.
3. Latency on the hot path AI inference is slow. Putting it on the synchronous request path means your user waits 8 seconds for a page load. Fix: async pipelines, background jobs, streaming where it matters.
4. No fallback The model provider goes down. Your system returns a 500. Fix: explicit fallback paths — cached results, degraded mode, a useful error — so the system degrades gracefully instead of crashing.
5. Prompt drift The prompt that worked in week one stops working in week six because the model was updated, the context changed, or the task evolved. Fix: version your prompts, regression-test outputs, treat prompts as code.
What actually closes the gap
The engineers who ship reliable AI systems do not have better ideas. They have better habits:
- They name failure modes before writing code
- They set hard budget ceilings before running anything
- They move AI work off the request path by default
- They build fallbacks before they need them
- They version and test prompts the same way they version and test code
This is what we teach in the programme. Not how to prompt — how to engineer.
The question is not “can I build an AI feature?” The question is “can I build one that is still running three months from now?”
If you want to answer yes to that, apply for the next cohort.