Hyacehila's Blog

What makes agents deliverable is not only the core loop, but the surrounding engineering that turns language requests into tool contracts, routing, validation, isolation, recovery, and governance.

Semantic embeddings are not only for retrieval. LSH filtering, Faiss deduplication, and K-center greedy sampling all use the same representation space for redundancy and coverage.

A survey of how rewards evolve from preference pairs, verifiable outcomes, process supervision, rubrics, and open-agent ranking, gradually redefining what a good answer means.

A technical evolution map for medical AI, from black-box predictors toward traceable medical agents.

Models have crossed the good-enough threshold. In 2026, the scarce resource is not the next bigger foundation model, but applications that enter workflows and daily life.

Reward writes goals into the optimizer, but it does not prove the model learned the right goals. This beginner-oriented rewrite explains why post-reward verification is needed.

SWE-bench, SWE-agent, and Tau-bench show that agent development cannot focus only on models and frameworks. Tasks, tools, users, state, rules, and validators shape whether agents work.

Following the signal chain from reward to baseline, advantage, and normalization, this post explains why RL algorithms for LLM alignment keep rewriting reward signals.

Better simulators often come from a clear division of labor: encode conservation, geometry, causality, boundaries, and rendering structure, then let learning cover expensive or fuzzy parts.

Agent Skills are popular because they package capabilities with low friction. They are a useful layer of context engineering, but not a new universal protocol or final agent form.