Building Reliable Agentic Systems
Reliability in agents is not a model problem. It is an architecture problem, and it is solved in the boring layers.
Reliability in agentic systems is usually discussed as a model property — a better model will hallucinate less, plan better, recover more gracefully. In practice, the systems that run every day are not the ones with the best model. They are the ones with the clearest seams.
Three seams that matter
Memory. An agent that cannot remember what it decided yesterday will re-derive it today, differently. Memory has to be a storage decision, not an emergent one.
Tools. A tool that fails silently teaches the model to keep going. Every tool we ship fails loudly, with a message the model can act on.
Evaluation. If you cannot measure a regression, you will ship one. Evaluation runs before deployment and on a fixed set, not on whatever the last demo used.
The uncomfortable part
Most of the work is not modelling. It is schema design, retry semantics, and deciding what the system should refuse to do. That work does not demo well, and it is the only reason anything survives contact with real use.