W31 - "Context Engineering for AI Agents"

Manus’s blog post “Context Engineering for AI Agents” gave me a more pragmatic understanding of AI application development.

In the past it was easy to reduce AI application development to simply “calling an API”; now I realize the real challenge is how to use traditional, deterministic software engineering methods to manage an inherently uncertain AI. Engineering experience accumulated through countless trials and errors is the technical barrier for handling large-model uncertainty. This barrier cannot be easily judged from code or product demos; only conversations with engineers who have deep experience can truly validate whether a systematic advantage exists.

The core competitive advantage in building applications is the ability to translate fuzzy human intentions into workflows that machines can execute reliably.

A book on complex systems, Discovered, Not Designed, gave me a more coherent understanding of Context Engineering — that elusive “art of deterministic orchestration.” The book divides problems into two types: difficult problems, which traditional engineering methods can explicitly design solutions for; and hard problems, which are often hard to explain and cannot be solved by explicit design, but are addressed through continuous trial, error, and self-adjustment.

Industrial products like airplanes are complex but not complex systems; they belong to the category of explicitly designed complexity. Large models, by contrast, are closer to complex systems — products that are “discovered.” When training large language models, we don’t directly modify the model’s internal parameters by hand; instead we use external feedback to guide the model’s self-adjustment until it produces the desired outputs. Similarly, market prices are not the result of careful human design but emerge spontaneously from the interactions and strategic behavior of countless traders.

Uncertainty is a blessing, not a curse. It’s important to understand the root of large-model uncertainty and place it in the right problem domain. We’re facing a brand-new engineering field that is still in its early stages, with a lot of messy, concrete work to do. We must not cling stubbornly to the deterministic mindset suited to difficult problems.

Last updated