Artificial intelligence agent development is moving away from prompt tuning and toward deterministic harness engineering to prevent systemic execution failures. Speaking at the AGNTCon conference and sharing his remarks on his website, Drew Breunig, chief executive of AI engineering firm Cmpnd and author of an O’Reilly handbook on context engineering, outlined the structural limits of relying purely on probabilistic language models. Breunig explained that despite context windows expanding to millions of tokens, real-world agents continue to suffer from performance degradation and errors. To ensure reliability, developers must rely on standard, deterministic software code to assemble, filter, and control data pipelines rather than leaving context assembly to stochastic model calls.
According to Breunig’s analysis, the application ecosystem has evolved from static prompt engineering in 2023 through dynamic context engineering in 2025, reaching harness engineering as the new standard in 2026. The previous approach of pouring massive volumes of unstructured raw data into expanding context windows and refining natural-language prompts has given way to code-level data filtering. Breunig stressed that the core task of modern application design is explicitly dividing the deterministic domain governed by code from the probabilistic domain handled by model inference. Context assembly itself must progressively shift into deterministic software logic to prevent agent derailment.
A central technical driver behind this architectural shift is context inertia, a failure mode where an agent clings to flawed early hypotheses or sub-goals and rejects corrective signals from external tools. As irrelevant context accumulates, models suffer from distraction, compounding errors and driving up operational costs. Breunig highlighted mitigations such as Anthropic’s Memento harness and intentional forgetting techniques from researchers at UC Berkeley, which systematically purge stale and biasing context through deterministic code pipelines. Across developer forums such as Hacker News, discussions increasingly emphasize that enterprise-grade reliability requires wrapping probabilistic models in rigorous software boundaries rather than relying on natural-language steering alone.