ARCHIVED · JULY 2026
July 2026
The Agentic Design Library as published in July 2026 — preserved exactly as it appeared on the homepage that month. Reading list of 6 essay-style pieces.
One Useful Thing
June 30, 2026
The Twilight of the Chatbots
Ethan Mollick
Mollick argues that AI can now finish longer and more complex tasks on its own, so work is moving from chatting with a model to handing work to agents. He contends that domain expertise, not technical background, decides who gets value from that delegation. The piece is a field report on how work changes as AI capability keeps growing exponentially.
delegationhuman-oversightfuture-of-workhamel.dev
June 29, 2026
"It's Hard to Eval" Is a Product Smell
Hamel Husain
Husain argues that a product that is hard to evaluate is usually badly designed, because the same opacity makes it hard for users to check its work. Using three case studies, including an AI data agent, he shows how surfacing supporting evidence and splitting output into reviewable units makes evals tractable. Good evaluation and good human oversight turn out to be the same design problem.
evaluationhuman-oversightproduct-designAhead of AI
June 27, 2026
Using Local Coding Agents
Sebastian Raschka
Raschka walks through building a fully local coding agent by serving open-weight models through Ollama and connecting them to harnesses such as Qwen-Code, Codex, and Claude Code. He compares how different harness and model pairings perform and flags the security review that agent codebases need. The piece makes the case that the harness is a separable, inspectable layer of agent architecture.
agent-harnesscoding-agentsopen-weight-modelseugeneyan.com
June 21, 2026
Patterns for Building Cybersecurity Evals
Eugene Yan
Yan reviews seven cybersecurity benchmarks, from capture-the-flag exercises to multi-host network attacks, and finds the same four parts in each: a sandboxed target, difficulty-setting inputs, agent tools, and a deterministic grader. He argues for partial credit through subtasks along the attack chain, so evaluators can see where an agent fails, not only whether it fails. The result is a reusable template for evaluating agents on long, multi-step tasks.
agent-evaluationbenchmarkssecuritySimon Willison's Weblog
May 6, 2026
Vibe coding and agentic engineering are getting closer than I'd like
Simon Willison
Willison admits that as coding agents become more reliable, he reviews less of the code they produce, which blurs the line between vibe coding and disciplined agentic engineering. He works through the tension between trusting proven tools and staying accountable for the result. He argues the bottleneck has moved upstream to design and requirements, and that agent-built projects should be judged by evidence of real use.
agent-assisted-engineeringaccountabilitycode-reviewhuyenchip.com
January 7, 2025
Agents
Chip Huyen
Huyen's long essay defines an agent by its environment and its tools, with the foundation model as planner. She argues that an agent succeeds or fails on two things, its tool inventory and the strength of its planner, and covers planning, reflection, and failure modes in depth. It is included as this issue's one foundational piece, the conceptual basis that the newer essays build on.
agent-architectureplanningtool-use