· 1 min · agents
the lethal trifecta audit takes ten minutes
Private data, untrusted content, a way to phone home. Your agent only needs all three at once to ruin you, and yours probably has all three.
i ship llm features for a living. i write down what breaks, what it costs, and what i'd do differently.
· 1 min · evaluations
temperature zero is not determinism
I pinned an eval to exact-match at temp 0 and it flaked anyway. Where the wobble actually comes from and what to assert instead.
· 4 min · architecture
"Global Standard" is a billing tier, not a failover plan
Azure OpenAI lost Sweden Central for six hours last week, then the gateways went in 18 regions. Microsoft's own docs say your global deployment dies with its primary region.
· 4 min · agents
your agent's "read-only" internet access was never read-only
Wikimedia says OpenAI agents hit it with millions of requests and tried to turn Etherpad into a proxy. The GET-only sandbox they came from is probably your sandbox too.
· 1 min · architecture
The batch API is 50% off and your cron jobs don't know
Both major providers sell the same tokens at half price for anything that can wait a few hours. The three-question test for what belongs there.
· 2 min · rag
Embeddings are not search
Cosine similarity finds things that vibe with your query, which is not the same as things that answer it. Why every RAG stack still needs a keyword path, with diagrams.
