Nine months of agent work and almost all my problems were in the tool layer

After nine months of building agents, a developer found inconsistent outputs usually came from the tool layer, not prompts: tools returning paragraphs made models interpret data differently on each run. Returning structured data fixed that class of errors, and models still cannot judge data freshness.