A Databricks benchmark showed a leaner Pi harness using roughly 3x less context per turn and delivering more than 2x lower task cost at comparable quality. The deeper enterprise implication is control: if every team builds its own agent wrapper with a different tool set, permission model, and safety prompt, you don't have a platform — you have a fleet of fragmented runtimes.

A Stanford study across 500+ agent tasks found agentic runs consume up to 3,500× more tokens than single-turn calls — often looping over bloated context logs without making progress. Three architectural patterns are essential to scale economically: context budgets before orchestration, queryable context instead of cumulative, and strict schema validation before inference.

Agent stacks keep rebuilding the sensor-fusion pipeline from scratch, without the pieces that made it work. Priors, calibrated gates, coordinate transforms, and recovery are not new ideas — but every one of them is missing from most agent architectures today.

Claude Code surfaces an interesting tradeoff between skills and hooks — but governance ends up spread across hooks, settings, plugins, and frontmatter rather than living natively in the workflow itself. The next evolution is runtime-native governance, where key checks become structural guarantees instead of distributed control points.

Learn how Box moved from early LLM wins to an agentic AI platform that handles complexity, boosts accuracy, and evolves with enterprise needs.