Pavamana AI Labs · The Hard 70%
The Hard 70%. From Pavamana AI Labs.
The hard 70% of shipping a reliable agent is the architecture, not the demo.
Try: tool misuse
AI Engineering
An AI agent ran a real café for two months. Here is what it could not learn.
Andon Labs opened a physical café in Stockholm and let an AI agent run it. Gemini 3.1 Pro lost around $6,000. GPT-5.5 stopped the losses and shrank the menu to coffee only. Neither model was wrong. Both were filling a gap that should have been filled by validated operating boundaries.
Balagei G Nagarajan12 min read
Failure Rates at 2 Agents vs. 5 Agents vs. 10 Agents: What the Data Shows
Balagei G Nagarajan4 min read
Why do agents declare the job done before verifying it worked?
Balagei G Nagarajan5 min read
Why do agents ignore inputs from peer and subordinate agents?
Balagei G Nagarajan5 min read
Why does role confusion get worse as you add more agents?
Balagei G Nagarajan5 min read
How does your agent topology — manager-worker or debate — change what breaks?
Balagei G Nagarajan5 min read
How can an AI agent be used as a proxy to attack connected systems?
Balagei G Nagarajan5 min read
How does "not invented here" thinking slow AI agent adoption?
Balagei G Nagarajan5 min read
Why do agents that work perfectly in narrow scopes create the most dangerous expansions?
Balagei G Nagarajan5 min read
Eighteen bugs that broke my automatic guest responder
Balagei G Nagarajan7 min read
A people-pleasing AI agent is a production liability
Balagei G Nagarajan7 min read
Your tool's output schema changed. Your agent didn't notice.
Balagei G Nagarajan6 min read
Get the next one in your inbox
How to architect AI agents that survive production — patterns, decisions, and the engineering that actually ships. No noise.
From the team building Pavamana AI Labs, the pre-build AI agent doctor.