ARCHIVED · JUNE 2026
June 2026
The Agentic Design Library as published in June 2026 — preserved exactly as it appeared on the homepage that month. Reading list of 9 essay-style pieces.
Anthropic Engineering
May 25, 2026
How we contain Claude across products
Max McGuinness, Mikaela Grace, Jiri De Jonghe, Jake Eaton, Abel Ribbink
Argues that as agents grow more capable, approval prompts and model-level defenses are not enough on their own; the dependable control is a hard limit on blast radius enforced by the environment. Walks through how sandboxes, virtual machines and network controls are fitted differently to claude.ai, Claude Code and Cowork according to each product's users and risk profile.
safetycontainmentbounded-authorityShreya Shankar (personal blog)
May 21, 2026
Exploring Agent-Assisted Qualitative Analysis
Shreya Shankar
Reports on experiments with agents doing qualitative coding under six supervision conditions. Agents handle the mechanical work well but tend to paraphrase rather than synthesize, settle on interpretations too early, and wear out their human supervisors with low-value checking. The research problem, Shankar argues, is designing systems where human taste and agent scale work together, not automating the analysis away.
human-oversightevaluationhuman-agent-collaborationarXiv
May 12, 2026
Autonomy and Agency in Agentic AI: Architectural Tactics for Regulated Contexts
Damir Safin, Dian Balta
Treats agency (what a system can do) and autonomy (how much it does without a human) as linked design dimensions that have to be set together, each with five operational levels. Proposes six architectural tactics, from checkpoints and escalation to tool fencing and write staging, for moving a deployment within that space while keeping responsibility, auditability and reversibility explicit.
governancehuman-oversightarchitecturePhilipp Schmid (personal blog)
May 5, 2026
How Agents Manage Other Agents: Four Subagents Patterns in 2026
Philipp Schmid
Sets out four patterns for coordinating subagents, ordered by how much control and communication they allow, from subagents called like functions to direct agent-to-agent messaging. The practical advice is to start with the simplest pattern and move up only when the coordination problem, and the model's capability, actually call for it.
multi-agentorchestrationarchitectureEugene Yan (personal blog)
May 3, 2026
How to Work and Compound with AI
Eugene Yan
Offers a working method for delegating to AI built on five principles: context as infrastructure, taste encoded as configuration, verification as the condition for autonomy, scale through progressively larger delegation, and closing the loop after each session. Because every finished artifact becomes context and every correction updates configuration, the gains compound, much as they do when onboarding a human teammate.
agent-assisted-engineeringcontextverificationAndrej Karpathy (bearblog)
April 30, 2026
Sequoia Ascent 2026 summary
Andrej Karpathy
Karpathy's written account of his 2026 talk: the unit of programming has moved from lines of code to delegated macro-actions, and the programmer's job is shifting toward orchestration and quality control. Capability jumps where environments are verifiable, so human judgment is still needed to decide what is worth building and to notice when an agent has gone past what it can reliably do.
agent-assisted-engineeringverifiabilitysoftware-3.0Addy Osmani (personal blog)
April 28, 2026
Long-running Agents
Addy Osmani
Looks at agents that keep making progress across many sessions and context windows, and finds that the major platforms have settled on similar designs: reasoning kept separate from execution sandboxes and durable session logs. Finite context, missing persistent memory and the lack of self-verification are handled through structured artifacts, checkpoints and a separate evaluation layer, not by waiting for better models.
memorylong-horizonstate-managementAddy Osmani (personal blog)
April 19, 2026
Agent Harness Engineering
Addy Osmani
Argues that a coding agent's quality depends less on the model than on the harness around it: system prompts, tools, context policies, hooks, sandboxes and feedback loops. Each agent failure should be treated as a lasting signal and fixed in the harness, so a decent model with a good harness can beat a stronger model with a poor one.
harnesstool-usereliabilityLalit Maganti (personal blog)
April 5, 2026
Eight years of wanting, three months of building with AI
Lalit Maganti
A detailed account of building syntaqlite, a SQLite parser and language server, mostly with coding agents. Agents were strong at verifiable, tedious implementation and harmful when used for design decisions that had no objective check. The project only worked once the author took back control of the architecture, concluding that AI is a force multiplier for implementation but a dangerous substitute for design.
agent-assisted-engineeringcoding-agentsdesign