The central argument is that useful AI agents usually need less infrastructure than engineers expect. Drawing on Picasso’s gradual simplification of a bull into a few recognizable lines, Mathew describes moving away from elaborate frameworks, multi-agent systems, memory architectures, and orchestration toward a simple, task-focused workspace. In his approach, a folder holds Markdown instructions and context, reusable skills, and access to the tools the job requires. A coding agent such as Claude, Codex, Gemini, or pi can help set up and refine that specialized agent; the goal is not to create intelligence from scratch, but to guide an existing model toward a particular task.
Building an agent starts by defining its job, success criteria, decision-making limits, needed skills and context, and the smallest suitable toolset. It can then be invoked interactively or headlessly from a script, schedule, webhook, or sandbox. Improvement comes from reviewing results and session transcripts, diagnosing failures, and turning reusable lessons into better instructions, examples, or tests. Benchmarks, or evals, help determine whether changes actually improve performance. More complex work can be divided among specialist agents in separate folders, coordinated by a root agent. A website-publishing workflow illustrates the method: an agent edits content, builds and deploys the site, and checks the live result, while human review feeds corrections back into its editorial guidance. The recommendation is to solve the basic case first and add machinery only when the work requires it.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.