Posts

Much of the early conversation around applied AI focused on prompt engineering: if the output was weak, the prompt needed improvement. Recent work in agent engineering suggests a broader shift. Prompt engineering still matters, yet it now functions as only one layer in a larger system. Context engineering determines what information the model sees, when it sees it, and what gets preserved across work. Harness engineering goes further still: it defines the runtime environment, verification loops, handoff artifacts, decomposition strategy, and coordination structures that allow an agent to work productively across long horizons rather than a single short session. Anthropic’s recent engineering posts trace this progression clearly, from context curation to initializer agents, structured progress files, parallel agent teams, planner-generator-evaluator loops, and decoupled execution architectures. This essay argues that the next generation of AI products will be differentiated less by clever prompts alone than by the systems built around model reasoning (Anthropic Engineering, 2025a; Anthropic Engineering, 2025b; Anthropic Engineering, 2026a; Anthropic Engineering, 2026b; Anthropic Engineering, 2026c). ...

As AI agents move from demo toys to production infrastructure, the question of how they connect to backend services has become a serious architectural decision. There are currently two dominant strategies: exposing capabilities through REST APIs paired with structured skill documents, and exposing capabilities through MCP (Model Context Protocol) servers. The popular framing — that REST is legacy and MCP is the future — oversimplifies the tradeoff. REST APIs, particularly when backed by machine-readable specifications like OpenAPI, offer strong contracts, broad compatibility, and mature tooling. MCP offers agent-native tool discovery, session-aware invocation, and push-based capability updates designed specifically for LLM tool use. Neither subsumes the other. This essay examines the two strategies across three dimensions — contract strength, orchestration control, and adaptability to code changes — and argues that the right choice depends on who your consumers are, what granularity they need, and how much backend volatility you expect them to absorb. ...

Drug-drug interactions (DDIs) occur when multiple drugs react with each other when taken together. They can lead to unintended side effects that may be harmful to patients. Developing an efficient and accurate computational model for DDI predictions is highly important to assist healthcare professionals in making better prescription decisions. Model Architecture The proposed GMPNN-CS++ model employs the Self-attention mechanism and a residual memory network after GMPNN-CS’s message-passing module to enhance the extracted representation of cross-substructure pairs within two interacting drug molecules. ...

Finalist Award — Top ~7% globally at the 2023 High School Mathematical Contest in Modeling. The Model We constructed a Dandelion Spread PDE Model (DSM) — a system of four coupled partial differential equations that model population densities of: Settled dandelion seeds Dandelion plants Puffballs Drifting seeds We corrected the Fisher model by multiplying a logistic term to obtain a logistic population growth that depicts the effect of intraspecific competition on the dandelion population. ...

Advanced to Final Defense — Top 26 among 900+ teams at the IMMC Greater China International Round. Approach Applied machine learning methods — Entropy Weight and K-Means clustering — to evaluate different development plans for a 5-square-kilometer land area. The model considered three key dimensions: Environmental impacts — ecological footprint and sustainability metrics Economic profits — projected returns and cost-benefit analysis Community utilities — social infrastructure and quality of life factors The optimization framework balanced these competing objectives to find the most effective land development strategy. ...

Advanced to International Round at the IMMC Greater China National Round. Approach Applied machine learning classification methods to classify lizards into 26 species: Logistic Regression — baseline linear classifier Decision Tree — interpretable tree-based model Random Forest — ensemble method for improved accuracy The models were trained on morphological and behavioral features to achieve reliable multi-class species identification. Download Full Paper (PDF)

Global Meritorious Winner — Top ~10% among 11,296 teams. The only high school team to achieve this on Problem B (among 728 teams). Methodology We first built an optimization model by subtracting the predicted economic profit using the modeled negative impacts, and conducted simulated annealing to obtain the value of each variable for the optimized policy. To account for changes in animal populations and propose an alternative policy that preserves these populations, we developed our animal population model using the logistic model framework. This model includes the additional mortality rate resulting from policies that permit hunting. ...