ARCHIVED · MAY 2026
May 2026
The Agentic Design Library as published in May 2026 — preserved exactly as it appeared on the homepage that month. Reading list of 8 essay-style pieces.
arXiv (EMNLP 2026)
April 17, 2026
A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle
Zehao Lin, Xixuan Hao, Renyu Fu, Shaobo Cui, Kai Chen, Chunyu Li, Zhiyu Li, Feiyu Xiong
The survey treats persistent, writable agent memory as an attack surface and organises the threats across six lifecycle phases, from write and store through share, forget and rollback. It proposes Verifiable Memory Governance, a set of mechanisms for keeping memory state auditable and recoverable, and argues that protection has to start at storage, not retrieval.
memoryagent securitygovernancemartinfowler.com
April 2, 2026
Harness engineering for coding agent users
Birgitta Böckeler
Böckeler sets out a vocabulary for the controls developers wrap around coding agents: feedforward guides that prevent mistakes and feedback sensors that let agents self-correct, each either computational or model-based. Her conclusion is that linters and type checks handle structure well, but justified confidence in an agent's functional behaviour still needs a coordinated system of controls, not any single tool.
agent-assisted engineeringreliabilityspec writingOne Useful Thing
March 31, 2026
Claude Dispatch and the Power of Interfaces
Ethan Mollick
Mollick argues that much of what feels like a capability ceiling is really an interface problem: the chat box overloads users, and less experienced workers most of all. He looks at three ways out (profession-specific tools, personal agents reached through familiar messaging channels, and interfaces the AI generates on demand) and suggests that better interfaces will make existing models feel substantially more capable.
human-agent interactionpersonal agentsinterface designHamel Husain
March 26, 2026
The Revenge of the Data Scientist
Hamel Husain
Husain argues that the foundation-model era makes classical data-science discipline more necessary, because evaluating stochastic AI systems is an experimental-design problem. He names five common evaluation failures (generic metrics, unverified LLM judges, weak experimental design, bad labels, over-automation) and matches each to an established data-science practice.
evaluationLLM-as-judgemethodologyAnthropic Engineering
March 25, 2026
How we built Claude Code auto mode: a safer way to skip permissions
John Hughes
A candid design account of replacing per-action human approval with a two-layer defence: a prompt-injection probe and a transcript classifier that decides which agent actions may proceed. The essay is explicit about the trade-off, reporting a low false-positive rate alongside a meaningful miss rate on overeager actions, and frames auto mode as safer than unrestricted permissions rather than a substitute for careful human review.
bounded authorityhuman oversightagent safetyAnthropic Engineering
March 24, 2026
Harness design for long-running application development
Prithvi Rajasekaran
Borrowing the generator/discriminator split from GANs, the piece shows that separating the agent that builds from the agent that judges produces markedly better long-running software work than asking a model to grade itself. It also observes that stronger models need less scaffolding, so harness design should be revisited as capabilities change rather than treated as fixed infrastructure.
multi-agent systemsagent architectureevaluationInterconnects
March 22, 2026
Lossy Self-Improvement
Nathan Lambert
Lambert argues that AI-assisted AI development is real but linear, not a runaway recursive loop. Narrow automation, Amdahl's-law limits on parallel agents, and human and organisational bottlenecks make self-improvement 'lossy': many agents sampling similar solutions add less than their numbers suggest.
multi-agent systemsscaling limitshuman oversightarXiv
March 17, 2026
Runtime Governance for AI Agents: Policies on Paths
Maurits Kaptein, Vassilis-Javed Khan, Andriy Podstavnychy
This position paper makes the agent's execution path the central object of governance, defining compliance policies as functions of agent identity, path history, the proposed action and organisational state. It shows that prompt instructions and static access controls are special cases of this view, and it lists the open problems in governing non-deterministic, path-dependent agents at runtime.
governancebounded authorityposition paper