How to Build Long-Horizon AI Agents — Mitch Troyanovsky, Basis

Building reliable, long-horizon autonomous agents requires moving beyond simple outcome-based evaluations toward rigorous process-driven "behavior specs." Mitch Troyanovsky, co-founder of Basis, explains that agents tasked with complex, multi-step professional work—such as end-to-end accounting—must maintain coherence over extended durations, a feat currently hindered by the lack of inherent short- or long-term memory in LLMs. To bridge this gap, developers must employ context engineering and structured ontologies that act as "runtime training data," ensuring agents follow established professional protocols rather than relying on unpredictable, non-deterministic reasoning. While technical advancements like reinforcement learning and model-level self-improvement are significant, the primary challenge remains defining high-quality signal for non-verifiable domains. Ultimately, the competitive advantage for agent-native companies lies in deep business integration and workflow ownership rather than proprietary technical tricks, as model capabilities will inevitably commoditize current harness-based solutions.