O'Reilly AI & ML Radar

As AI coding agents become extremely proficient, the bottleneck has shifted from writing code to reviewing it. Data shows a dramatic increase in code churn, defects, and review time. The key is to adapt review processes based on the context: blast radius, code longevity, and team size. Capturing agent reasoning can alleviate review burden.

This article explores the 'U-shaped' context loss problem in LLMs, where models tend to ignore information in the middle of their context window. The author discusses recent research and presents five practical techniques to mitigate the issue, based on real-world experiences developing an AI-driven quality engineering skill.

The article argues that the AI agent industry is falling into a 'tool trap' by obsessing over protocols like MCP and AI Skills while neglecting the true strategic discipline: Agent Experience (AX). The author contends that protocols will keep changing, and understanding how agents interact with your systems and optimizing that experience is the key to long-term competitiveness. The piece outlines five steps to build an AX practice and emphasizes that AX is an extension of UX, DX, and CX.

This article introduces the concept of "principal drift" in enterprise agent architectures, where the human authority behind agent actions becomes increasingly detached as agents multiply and compose. It describes the cascade of identity collapse, authority erosion, and accountability dissolution, and proposes solutions including reasoning-grade audit records and a new "agent operations" function.

Loop engineering replaces direct prompting of coding agents with a system that recursively iterates toward a goal. It consists of five core components: automations, worktrees, skills, plugins/connectors, and subagents, plus external memory. Tools like Codex and Claude Code are converging on similar loop primitives, and subagents split ideation and verification to improve reliability.

This week, egghead.io cofounder John Lindquist joined host YK Sugi to discuss the contested release of Claude Fable 5, financial shifts in AI spending, and a practical framework for building in the agent era. Key topics include the government directive that pulled Fable 5, Uber burning through its 2026 AI budget by April, and John's 'ingredients beat inference' approach using open-source clones. The episode also covers the SpaceX acquisition of Cursor and Salesforce's purchase of Fin.

Kubernetes has evolved from a container orchestrator to a de facto AI platform, with 82% of container users using it in production in 2025. Generative AI and agentic AI workloads are increasingly run on Kubernetes, as highlighted by the CNCF survey and industry examples. Networking remains a fundamental skill gap, addressed by a new CNCF certification.

Many enterprises underestimate the complexity and long-term cost of building their own AI agent platform. This article analyzes four underestimated components—memory, governance, eval, and orchestration—and provides five questions to ask before committing to such a project.

Coding agents simplify building AI agent workflows but obscure nonlinear cost scaling. Classical CS optimizations like memoization, pruning, and dynamic programming are essential to avoid repeated work and high costs.

Agentic coding tools like Claude Code, Cursor, and Codex generate code that may be uncopyrightable, owned by your employer, or contaminated by open source licenses you cannot see. Some of this is settled law, some is actively contested, and this piece is clear about which is which.

This week Miguel Fierro, a former Microsoft principal researcher who recently founded his own company, RecoMind, joined data and AI evangelist Christina Stathopoulos to talk about the state of recommendation systems. Christina also ran through the latest AI news she’s been watching, from Anthropic’s continued rise to responsible AI, announcements from Google’s I/O 2026 conference, and (continuing the discussion from last week) the growing backlash against tokenmaxxing as a productivity metric. Here are three takeaways from the conversation.

This is the eighth article in a series on agentic engineering and AI-driven development. It addresses context loss in AI agents performing complex multistep tasks. The author introduces the Externalize-Recognize-Rehydrate (ERR) pattern: saving agent state to disk, detecting context degradation, and recovering from files. Historical analogies (640K memory limit) and a real Copilot crash example illustrate the problem. The article details externalizing two layers of state: execution continuity (current step) and task continuity (overall goals).

This article addresses common pitfalls from AI demo to production, offering a practical playbook covering latency budgets, fallback design, quality measurement, A/B testing, model drift monitoring, evaluation frameworks, graceful degradation, and prompt engineering.

GitHub Copilot's shift to usage-based billing on June 1 exposed the true cost of agentic workflows. This article analyzes token consumption, tool design impact, and strategies for prompt optimization and output formatting, emphasizing that cost control should be a platform governance issue.

This article updates the 2024 AI agents stack diagram, introducing a six-layer architecture for 2026: Models & Inference, Protocols & Tools, Memory & Knowledge, Frameworks & SDKs, Eval & Observability, and more. Key changes include MCP standardization, reasoning models, and memory as a first-class primitive. It offers honest takes and guidance on evaluating each layer.

On this week's episode, host Andreas Welsch and guests Maya Mikhailov and Doug Shannon discuss OpenAI's move into personal finance, metacognition as a professional skill, the backlash against token-based productivity metrics, and the limitations of forward-deployed engineers. The core theme: the AI industry is good at generating output but still figuring out what output is valuable.

DJ Patil's listening tour reveals a broken promise in AI, with students and workers feeling terrified. He proposes community makerspaces and emphasizes organizational capacity as the bottleneck. Data infrastructure is a competitive advantage, enabling companies like Devoted Health to leverage AI quickly.

Anthropic's recommendation to use EPSS for vulnerability prioritization marks a shift from static severity scores to predictive models. The article explores the machine-scale problem of vulnerability volume, the distinction between pointing machines and knowing machines, and the policy changes needed for security programs to survive the coming wave of AI-discovered vulnerabilities.

As syntax becomes cheap and abundant, architectural control becomes the scarce resource. Effective governance starts upstream, where intent, constraints, and threat models shape the agent’s working context before generation begins. The goal isn’t better prompting but build-time boundaries that prevent structurally invalid code from entering the system.

The article examines the growing trend of nations seeking technological sovereignty, using Brazil's pursuit of medical sovereignty as an analogy for AI. It argues that decoupling is too narrow a frame; instead, countries want to stay connected while building their own capacities, similar to federation rather than separation. Open-source AI models and protocols are key tools, but infrastructure (data centers, chips, power grids) is the critical layer that is hard to replicate. The piece envisions a federated AI future and the need to rebuild infrastructure for the AI era.

Despite the rise of AI agents leading many to declare the end of SaaS, this article argues SaaS is not dead. Work is collaborative, while agent-based programming is individualistic, lacking sharing, collaboration, testing, versioning, and security. SaaS companies can adapt by providing APIs for agents, becoming the system of record for data.

The article discusses the limitations of open-weight AI models and open protocols as open source strategies, using Anthropic's acquisition of Stainless as a case study to illustrate complement capture and moat migration in AI infrastructure. It argues that the developer experience layer is being consolidated by platform giants, creating new competitive advantages, and emphasizes the need to analyze dependencies within the ecosystem to identify potential chokepoints.

This article is the seventh in a series on agentic engineering and AI-driven development, focusing on context management in AI sessions. The author shares a personal experience with Gemini forgetting earlier notes, introduces the concept of context compaction, and provides four practical techniques: split discovery from documentation, use handoff documents, give acceptance criteria rather than procedures, and use spec documents as bridges. These techniques apply to both developers and regular users, helping reduce frustration caused by AI forgetting.

O'Reilly's Infrastructure & Ops superstream explored the infrastructure needs, costs, and security challenges of AI workloads. DORA's report shows AI increases code delivery by about 10% but reduces stability, adding verification costs. Experts emphasize platform engineering, governance, and cognitive debt, recommending investment in internal platforms to ensure production readiness for AI applications.

AI coding agents default to the shortest path to 'done,' skipping specs, tests, and reviews that senior engineers know are essential. Addy Osmani's Agent Skills project builds senior-engineer scaffolding for agents, using workflows instead of prose. It includes 20 skills across six SDLC phases, incorporating Google engineering practices. Key principles: process over prose, anti-rationalization tables, nonnegotiable verification, progressive disclosure, and scope discipline. The article also covers three usage modes and patterns to steal even without installing.

AI agents delegate tasks across systems, but current architectures lack authorization models for these delegation chains, creating security gaps like ghost permissions and broken audit trails.

For over a century, corporate prestige and budgets have been measured by headcount. In the AI era, this model is obsolete. This article introduces the concept of the agentic P&L, shifting from headcount empires to federated nervous systems with new metrics like contextual density, agentic throughput, and decision provenance. Using a compliance department in a Tier-1 bank as an example, it outlines the transformation path and the new organizational unit: the 3+N squad.

Current production agents lack identity, context persistence, and platform support, creating governance and reliability gaps. This article proposes four architectural bets: agents need identities, universal context, durable execution, and platforms.

PocketOS founder Jeremy Crane used Claude for routine DB maintenance, but the AI agent accidentally deleted the production database and all backups. Railway managed to recover the data. The incident highlights systemic weaknesses: overly broad tokens, non-expiring credentials, and lack of sandboxing. AI amplifies existing issues rather than being the root cause. Lessons include principle of least privilege, credential expiry, human-in-the-loop, and world models.

Companies are racing to leverage AI for productivity, but most pilot projects fail. Investing in open standards like Agent Skills, MCP, and plugins protects against vendor lock-in and lowers switching costs. AI artifact catalogs help organizations turn individual wins into shareable institutional knowledge, enabling cross-team and agent reuse.