20 entries
This article demonstrates how Jev, a typed decision-making model from TypeSafe, can evaluate agent responses by turning natural language judgment into structured classification with confidence scores. Rather than parsing generated prose, Jev returns constrained answers with probability and confidence metadata that lets developers identify uncertain evaluations and set appropriate automation thresholds based on risk tolerance.
Meta's engineering team describes an AI agent system that captures organizational expertise as structured, auditable knowledge separate from reasoning procedures. The system combines domain-specific knowledge architecture with a self-improvement loop that implements expert feedback without model retraining, creating persistent institutional memory that scales specialist knowledge across organizations.
Maryam Miradi presents a seven-step architecture for building closed-loop evaluations of multimodal AI agents at scale, specifically addressing the core challenge that AI can improve image quality while altering truth. The framework emphasizes faithful reproduction of content (matching ingredients, portions, and identity) through human ground truth anchors, multimodal routing, cross-modal verification, QA gates with retry logic, production drift monitoring, and iterative tuning before deployment.
DAIR.AI's weekly survey synthesizes four key AI research directions: self-improving agents that evolve their own scaffolds and weights, metacognition frameworks that unify confidence and self-regulation in LLMs, the structural conditions required for meaningful routing in multi-model systems, and empirical critiques of harness evolution showing that simple search baselines often outperform supposedly optimized prompts. The curated selection positions these papers as navigational tools for builders implementing agent systems in production.
The article argues that the most effective workflow for creating skill.md files (structured instruction files for AI agents) combines AI-generated drafts with human refinement. It contrasts purely automated generation, which lacks nuance and product-specific judgment, with fully manual creation, which is time-consuming. The recommended hybrid approach leverages AI for structure and speed while relying on human expertise to add the contextual knowledge that makes AI agents actually effective.
Watching Maggie Appleton's GitHub Universe talk on agent collaboration. Running notes on alignment as the new bottleneck, the ACE environment, conversations under the hood, and a comparison with Engelbart's NIC.
This article presents a comprehensive architectural framework for agentic AI systems in 2026, arguing that most AI failures stem from architectural problems rather than model quality. The author defines four canonical design patterns (Reflection, Tool Use, Planning, and Multi-Agent) and emphasizes that agentic AI represents a paradigm shift from monolithic systems to distributed, observable, and bounded agent architectures.
This article argues that 2026 marks a fundamental shift from generative AI (passive, read-only text generation) to agentic AI (active, autonomous systems that plan and execute work). The author contends that AI engineers must move beyond prompt engineering and embrace systems engineering, treating AI as operational infrastructure that performs end-to-end tasks rather than conversational oracles that simply generate responses.
A DeepLearning.AI short course taught by Atai Barkai (CopilotKit co-founder) covering three patterns for letting AI agents render dynamic interfaces instead of text-only replies: controlled components, the declarative A2UI framework, and open-ended UI via MCP. The capstone builds a LangChain plus React plus CopilotKit agent that produces charts and interactive widgets.
This is a documentation hub for GitHub Copilot adoption and best practices at NAV (the Norwegian Labour and Welfare Administration). The site provides news updates, tools, guidelines, and usage statistics for AI-driven development, focusing on custom agents, skills, and organizational implementation of GitHub Copilot across teams.
Capturing ideas from the Guildford session and shaping them into a talk.
On working with Claude: unlike traditional software, LLMs have no pre-baked concept of operations. You have to define it yourself, and that requires an expert-level mental model.
Microsoft researchers introduce DELEGATE-52, a benchmark of long document-editing workflows across 52 professional domains. Even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupt about 25% of document content by the end of these workflows, and agentic tool use does not improve performance. The paper argues current LLMs are unreliable delegates: they introduce sparse but severe errors that compound silently over long interactions.
Issue of the Convocat newsletter Creative Prompts: garden updates, news for content professionals, seeds, and what I'm reading.
I rebuilt Karpathy's LLM wiki on top of my digital garden content, with interesting results.
6 plugins that turn Claude Code into a full dev team in 5 minutes. Each one covers a different role; planning, design, code review, security, memory, and team coordination.
Anthropic's project to secure critical software for the AI era, and why AI slop might be the bigger threat.
Five short experiments validating new Claude features (March 2026 release) on the digital garden. Practical demos for conversation designers.
Testing Claude Code's new scheduled tasks feature by automating weekly garden health checks.
A collection of thoughts & ideas on alternative metaphors for human-GenAI-interaction.