ARCHIVED · AUGUST 2026
August 2026
The Agentic Design Library as published in August 2026 — preserved exactly as it appeared on the homepage that month. Reading list of 7 essay-style pieces.
Addy Osmani
July 15, 2026
Own the Outer Loop
Addy Osmani
Osmani separates the agent's inner loop (investigate, implement, verify) from the human's outer loop (set constraints, review evidence, deliver the verdict) and argues that engineers must keep accountability for the outer loop. He names three hidden costs of agentic work, namely cognitive surrender, cognitive debt and orchestration tax, and makes answerability the basis for scaling it responsibly.
accountabilityhuman-oversightagent-assisted-engineeringParlance Labs
July 11, 2026
Do Automated Evals Work?
Antaripa Saha and Hamel Husain
The authors test three agentic evaluation systems against human-annotated traces from a production leasing assistant. The best recovered 87% of human-identified failures and surfaced some that humans missed. All three missed subtle failures where the exchange looked correct but served the user poorly, which leads the authors to recommend keeping humans in an iterative annotation loop rather than fully automating error analysis.
agent-evaluationhuman-in-the-loopllm-as-judgeLil'Log
July 4, 2026
Harness Engineering for Self-Improvement
Lilian Weng
Weng argues that the harness around a base model (workflow design, context management, tool integration, persistent memory) is a primary lever for capability, not a thin wrapper. She traces a progression from hand-built harnesses to systems in which coding agents optimise their own surrounding structure, framing this as a route to self-improvement that does not touch the weights.
agent-architecturememoryself-improvementArmin Ronacher's Thoughts and Writings
July 4, 2026
Better Models: Worse Tools
Armin Ronacher
Ronacher reports that newer frontier models make more malformed tool calls against third-party schemas, adding invented fields even when the core arguments are correct. He traces this to post-training on a forgiving first-party harness that quietly repairs bad calls, and shows that strict, grammar-constrained tool invocation removes the failures. The piece is a concrete warning about harness overfitting and tool-call reliability.
tool-usereliabilityharnessAddy Osmani
July 2, 2026
Agentic Autonomy Levels
Addy Osmani
Osmani replaces the single autonomy scale with two axes, agency and orchestration, and sets out six levels running from assisted suggestions to agent factories with minimal human intervention. His core rule is to match autonomy to risk, using three questions: how quickly we will know we are wrong, how cleanly we can undo it, and what proves we are right.
bounded-authoritygovernanceorchestrationGeoffrey Litt
July 2, 2026
Understanding is the new bottleneck
Geoffrey Litt
Litt argues that as agents get better at checking their own work, the reason for humans to understand code shifts from verifying it to taking part in shaping it. He borrows three techniques from education (explainer documents with quizzes, interactive micro-worlds, shared team spaces) to counter cognitive debt and keep people deeper in the loop rather than removed from it.
agent-assisted-engineeringhuman-oversightcognitive-debtGoogle DeepMind
June 18, 2026
Securing the future of AI agents
Rohin Shah and Four Flynn
DeepMind sets out an AI control roadmap that treats capable agents as potential insider threats and applies defense-in-depth from cybersecurity: trusted monitors, detecting misaligned actions before they cause harm, and safeguards that scale with capability. The piece draws on analysis of a million coding-agent trajectories and adds a three-layer view of agent security covering individual agents, multi-agent systems and the wider ecosystem.
safetyai-controlmulti-agent