Coding Agents and Skills
This page tracks coding agents, agent skills, and optimization methods for agent behavior. It connects directly to agent-skills and llm-wikis.
Sources in this batch
- Sebastian Raschka’s “Components of a Coding Agent” breaks down coding-agent architecture.
- Simon Willison’s LLM shebang note treats an LLM as an executable script interpreter.
- Wix reports 250 agent evals testing whether skills beat docs, with a nuanced outcome rather than a simple win.
- AlphaEvolve frames a coding agent as a tool for scientific and algorithmic discovery.
- Unsloth documents running local LLMs with Claude Code; Anthropic’s Claude Cowork extends Claude Code-like workflows to knowledge work.
- GEPA proposes a universal API for optimizing text parameters.
- Hermes Agent, Anthropic Courses, Open WebUI Open Terminal, and DSPy tutorials are part of the practical agent/tooling ecosystem.
Research interest
The surprising research thread is that agent capability may be less about raw model intelligence and more about the external scaffolding: skills, executable environments, eval loops, optimizer feedback, and local tool access. The Wix eval result is especially useful because it questions the naive claim that “skills beat docs” uniformly. GEPA and DSPy point toward a broader view: prompts, instructions, and agent policies can be optimized as first-class artifacts, not just hand-written.
Open questions:
- Which skills transfer across tasks, environments, and models?
- Can agent scaffolds be evaluated independently of the base model?
- How much of coding-agent progress comes from better model priors versus better runtime loops and feedback?
- Can text-parameter optimizers replace prompt-engineering folklore with reproducible search?