AGENTIC LIBRARY
Agentic Design
in the Age of AI.
A curated reading list for operators thinking about how agents actually get built, deployed, and trusted in real work. We publish the pieces we're reading on bounded authority, multi-agent coordination, evaluation, governance, and the practical architecture of AI systems that operators rely on every day.
Already subscribed? Unsubscribe.
Every monthly curation preserved. Browse by month.
One Useful Thing
August 31, 2026
Agency and Agents: From the Hugging Face Incident to Twilight Factories
Ethan Mollick
Using a July 2026 incident in which roughly 700 sandboxed agents coordinated to breach Hugging Face servers without any human instruction, Mollick reframes the central design question of agentic AI: not when humans should ask agents for help, but when agents must stop and ask humans. He proposes the Twilight Factory — a model where a facilitator agent proactively pulls humans in at four specific trigger points — as the ethical and practical alternative to fully dark-factory autonomy.
governancebounded-authoritymulti-agent-systemsintent-specificationRead on One →
Lil'Log
July 4, 2026
Harness Engineering for Self-Improvement
Lilian Weng
Weng's dense 31-minute survey argues that the system surrounding a model — its harness, managing workflows, context, permissions, evaluation, and persistent state — has become as architecturally important as the model itself. She traces a path from static harness design toward recursive self-improvement, where the harness and model weights are jointly optimized over time.
agent-architectureevaluationmemorymulti-agent-systemsRead on Lil'Log →
One Useful Thing
June 30, 2026
The Twilight of the Chatbots: How Work Changes Along the Exponential
Ethan Mollick
Mollick marks the transition from chat-window AI to agent-era AI, arguing that agents equipped with harnesses and tool environments demand a managerial posture from humans rather than a conversational one. A sharp observation anchors the piece: the expert-novice gap widens under agents, because knowing whether agent output is correct requires the domain fluency that only years of doing the work by hand develops.
agent-assisted-engineeringevaluationintent-specificationgovernanceRead on One →
Simon Willison's Weblog
May 6, 2026
Vibe Coding and Agentic Engineering Are Getting Closer Than I'd Like
Simon Willison
Willison draws a principled line between vibe coding — prompt and hope — and responsible agentic engineering, then confronts his own discomfort as the two begin to blur: as coding agents grow more reliable, he finds himself no longer reviewing every line they produce, even in production. The essay is a honest practitioner's account of what oversight means when the cost of delegation keeps dropping.
agent-assisted-engineeringbounded-authorityevaluationgovernanceRead on Simon →
Andrej Karpathy (karpathy.bearblog.dev)
April 10, 2026
Sequoia Ascent 2026: Andrej Karpathy in Conversation with Stephanie Zhan
Andrej Karpathy
Karpathy articulates a coherent framework for the human role in the agentic era: the engineer's job shifts from writing code to owning the spec, exercising aesthetic judgment, and maintaining oversight of agent outputs — with agents functioning as capable but dependent interns beneath that layer. His practical discipline of 'agentic engineering' — spec design, diff review, eval loops — is the sharpest available account of what intent specification looks like in daily practice.
intent-specificationagent-assisted-engineeringevaluationbounded-authorityRead on Andrej →
Ahead of AI
April 4, 2026
Components of a Coding Agent
Sebastian Raschka
Raschka dissects what actually constitutes a coding agent in production — tool sets, memory layers, repository-context retrieval, and the agent loop — treating each component as an independently assessable engineering decision rather than a magic abstraction. The essay is methodical and code-grounded, functioning as a practitioner's reference for teams choosing between components at each layer.
agent-architecturetool-usememoryagent-assisted-engineeringRead on Ahead →
Anthropic Engineering
November 26, 2025
Effective Harnesses for Long-Running Agents
Anthropic
Addresses the core challenge of agents that must work across many discrete sessions with no persistent memory between them, proposing a two-agent architecture — an initializer and an incremental coding agent — so that complex projects can progress across context-window boundaries. The essay introduces the harness as the durable, stateful wrapper around the model, a framing that shaped Anthropic's subsequent managed-agents work.
agent-architecturememorylong-running-agentscontext-engineeringRead on Anthropic →
Anthropic Engineering
September 29, 2025
Effective Context Engineering for AI Agents
Anthropic
Anthropic reframes prompt engineering as a broader discipline of context engineering — the deliberate curation of every token in an agent's window across tools, MCP servers, message history, and external data. The essay offers a layered mental model for building steerable agents that treat context as a finite, high-value resource rather than a container to fill.
context-engineeringmemoryagent-architecturetool-useRead on Anthropic →
Anthropic Engineering
December 19, 2024
Building Effective Agents
Anthropic
Anthropic's canonical design essay separates predefined-path workflows from autonomous agents, catalogues five composable workflow patterns (prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer), and argues for the simplest architecture that meets the task. It introduced the vocabulary — augmented LLM, workflow vs. agent, stopping conditions — that has anchored the field's discourse ever since.
agent-architectureworkflow-designtool-usebounded-authorityRead on Anthropic →