Mark Lummus and Navinkumar Patil explain why agents can struggle with interactive command-line interfaces: terminal redraws and lost state can make a visible human prompt effectively invisible to the model. Flags remain useful for simple independent options, but contextual questions and environment-dependent choices need a richer interface.
Mark Lummus and Navinkumar Patil use a coffee-ordering workshop to compare three approaches: controlling a terminal session with tmux, exposing machine-readable state and input through side channels, and providing the agent with relevant context. They encourage designs that work for both human and machine readers without discarding interactive guardrails.
Mark Lummus and Navinkumar Patil discuss the value of fast, deterministic tools beneath a reasoning agent. Participant results also reveal limitations of a latency-based benchmark: a narrowly optimized parser can score well without demonstrating the broader intended workflow. The practical lesson is to make tool behavior explicit and evaluate the task that actually matters.
Watch on YouTube
