Reading

In the video "Let Your Agent Cook: Using Skills to Evaluate and Improve Your App" published on the AI Engineer channel on October 11, 2026, Ankur Duggal, an engineer at Arize AI, demonstrated a workflow for diagnosing and improving LLM applications using coding agents. The video features a financial agent that generates seemingly reasonable but actually context-lacking investment advice, with its own evaluation system flagging issues such as wrong summaries, repeated tool calls, and fabricated answers.

Duggal showed how to connect these failure phenomena to a coding agent via Arize's Command Line Interface (CLI) and skills. The coding agent was tasked with investigating traces to identify specific interactions between supervisor agents, specialist agents, model calls, and tools. This process transformed vague quality complaints into evidence about exact failed steps. For instance, in an earlier experiment targeting stock queries outside the United States, Duggal pointed out that changing the country context alone was insufficient; the underlying tool also had to be updated to resolve the issue.

The video also discussed leveraging dataset, evaluator, and experiment skills to make proposed fixes testable rather than performing a series of unverified edits. Duggal addressed questions regarding local experiment execution, judge model selection, evaluation costs, sampling strategies, and the risk that a coding agent might spend effort on linting instead of addressing actual defects. At the close of the video, the live repair was still ongoing.

For developers seeking to improve AI agent reliability, this resource offers a concrete operational perspective. Duggal recommended human review of proposed changes and tailoring evaluation systems to specific application failure modes, ensuring that performance regressions can be detected promptly as traffic patterns and base models evolve.

According to reports by ZDNet Korea and Yonhap News, the office of Son Myung-soo, a member of the National Assembly's Standing Committee on Government Operations from the Democratic Party of Korea, recently analyzed data from the Ministry of Land, Infrastructure and Transport's Apartment Management Information System (K-apt) and financial institutions. The results show that in South Korea's annual apartment management fee market, valued at over 33 trillion won (approximately $23 billion), a company referred to as 'Company A' holds an overwhelming market share.

As of September this year, Company A accounted for 71.4% of the ERP (Enterprise Resource Planning) market based on household numbers and 68.2% based on apartment complexes. These systems handle fee calculation, accounting processing, and billing management. In comparison, the second-ranked competitor holds only 11.5%, leaving a gap of nearly 60 percentage points. This means more than seven out of every ten apartment households in South Korea use the same solution. Switching costs are high due to the need for historical financial data migration, system integration, and retraining of management personnel, leading to deep user lock-in.

Even more severe concentration exists in the Electronic Bill Presentment and Payment (EBPP) infrastructure connected to financial institutions. Among the major four banks and six card companies where linked operators could be identified, Company A's share of management fee payment integrations reached 99.7% to 100%. Last year, these financial institutions processed a total payment volume exceeding 11.47 trillion won through Company A, with 2.512 trillion won via banks and 8.9615 trillion won via credit cards.

This structure, where a single vendor controls both core management software and financial settlement channels, significantly raises barriers to market entry. Even if emerging IT startups introduce AI-based management fee analysis or accounting automation tools, they struggle to gain sufficient market access. The lawmaker pointed out that lack of competition may lead to degraded service quality and poses risks of widespread business disruption during system failures.

Son Myung-soo called on the Fair Trade Commission of South Korea to closely examine this de facto monopoly to establish a fair competitive environment. This finding highlights how underlying technical infrastructure can form invisible monopolies in seemingly fragmented residential service markets, suppressing innovation.

In its Form S-1 filed with the U.S. Securities and Exchange Commission (SEC), American AI company Anthropic disclosed an unusual founder succession plan. Under this scheme, each co-founder can nominate a ranked list of up to 10 candidates who would inherit their voting power if the founder becomes incapacitated or passes away.

This mechanism aims to ensure that control over the company persists even as founders sell shares or depart. The list remains confidential, and the order of succession is fixed in advance. Any nominee requires approval from at least five of the other six co-founders. While the board's governance committee can demand reconsideration, it cannot veto an approved list.

This control is enacted through five types of shares (A, E, F, N, and T). Public investors can only buy Class A shares, while Class F shares grant the founders 50.1% of voting power, separating economic ownership from control. Previous reports noted that the 'Founder LLC' entity holds this 50.1% voting stake, but the newly disclosed succession details reveal how this structure operates specifically at the founder level.

A data pipeline stripped minus signs from loan amounts, turning a net $784,000 into $5.3 million in debt, with multiple models validating this incorrect result. Hrishi Olickel, founder of Southbridge AI, used this failure at the October 2026 AI Engineer conference to illustrate his core argument: the final bottleneck in data transformation—intelligence and judgment—is now cheaply accessible, but the true difficulty lies in not knowing what information means until a decision is made.

Olickel posits that every data record is constrained by four modes. Using temperature as an example: the source mode determines whether you have a regional phone reading, a desk sensor measurement, or a watch reading; the storage mode limits what can be recorded—device firmware may be omitted, and recording a timestamp to the second might take longer than writing the record itself; the decision mode determines which omissions are fatal—assessing comfort requires multiple measurements throughout the day and confirmation of the measurer, while weather mapping cares about location and device changes, and installing sensitive equipment requires excluding the person as noise. The fourth mode is the model's prior, its belief about how the world 'should be.' These four modes are only simultaneously known when the decision actually occurs. He cited the term 'customer': for billing, it is the payer; for support, the caller; for compliance, the entity the organization is responsible for. A unified customer box cannot resolve these distinctions before the question is asked.

His alternative is a 'just-in-time ontology': building transformation structures when the decision is known, preserving their context and evidence for reuse. Operationally, he splits data work into three model roles. Designers decompose problems, write specifications, and authorize budgets, suited to slower, heavier reasoning models. Executors read PDFs, write database records, and check results within pipelines, suited to cheap, fast models with predictable defects. Optimizers start from a working run, try alternatives, debug, take notes, and adjust budgets to limit exposure, requiring reusable knowledge and stable execution environments. This separation explains why a new executor model could enter the Ohio pipeline without taking over its design—the system can test where the model helps while leaving specification and optimization to models suited for those tasks.

The Ohio campaign finance entity resolution project provided concrete figures. Olickel reported that the project was originally expected to cost about $10,000; after introducing a new decision model, the team completed all donor cases in ten days for under $800. The statewide job cost $795, with the first answer costing $128, while second opinions, disagreement resolution, and auditing totaled $545—checking costs were roughly four times the answering cost. On approximately 200 held-out families, a cheap reasoning model alone achieved results comparable to a comparison model at about 2% of its cost. A key finding was that a decision-first approach correctly answered core questions but missed nine household links because it confidently answered 'no' to each association. The team did not change models or thresholds but rewrote the household question to clarify 'what counts as sufficient evidence,' causing thirteen development questions to shift from no to yes. This demonstrated that definitions themselves can change results without changing the model.

OpenAI released a video case study in October 2026 focusing on the distribution strategy of Stack Overflow's latest product, 'Stack Overflow for Agents.' The company is building a knowledge intelligence platform designed to turn information scattered across workplace tools into verified knowledge for both people and AI agents.

According to the video description, the new product was built from the ground up by two engineers in just a few weeks using Codex. To distribute it, Stack Overflow launched a ChatGPT plugin. The company claims that this move caused signups to triple overnight, while users also created up to 10 times more agents.

The video demonstrates how the team accelerated development and brought company knowledge into tools people already use. For readers interested in AI industry investment, this case offers a direct data point on leveraging mainstream AI interfaces like ChatGPT as distribution channels to rapidly boost early-stage product adoption rates, although these specific metrics remain vendor-self-reported.

South Korean document AI company Epapyrus announced on October 27 that its open-source document processing library, PyMuPDF, has been updated to version 2.0. The core change in this release is the consolidation of three previously independent packages—the base engine PyMuPDF, the Large Language Model (LLM) data conversion tool PyMuPDF4LLM, and the layout analysis engine PyMuPDF Layout—into a single PyMuPDF package. Developers can now obtain all functionalities with a single pip install pymupdf command, eliminating the need to manage separate versions or adhere to specific import orders.

The new version includes built-in Graph Neural Network (GNN)-based layout analysis, capable of accurately identifying reading order in multi-column documents, footnotes, and tables, and converting extracted results directly into Markdown or JSON formats suitable for LLM input. Epapyrus stated that the process runs entirely on CPUs, requiring no GPUs or cloud calls. Compared to solutions relying on Vision-Language Models (VLMs), hardware resource consumption is approximately one-hundredth, with speed improvements of up to ten times. This makes it suitable for local or closed-network environments with strict data privacy requirements.

For existing users, PyMuPDF 2.0 maintains backward compatibility. Legacy import pymupdf4llm statements continue to work via a compatibility layer, and original MuPDF engine functions, license policies, and rendering/editing capabilities remain unchanged. According to Vice President Kim Jung-ah, PyMuPDF's global cumulative downloads have exceeded 1.2 billion.

Irving Tan, Chief Executive Officer of US storage device manufacturer Western Digital, recently participated in an interview on CNBC's financial program "The Exchange," engaging in a deep discussion with host Kelly Evans about the company's financial performance in the first quarter of fiscal year 2027 and its views on the artificial intelligence infrastructure market. This interview took place following the release of the company's strong earnings report, providing investors and industry observers with a first-hand interpretation of current market dynamics from management.

According to the episode description, Tan focused on analyzing trends in AI-driven storage demand. He posited that AI is fundamentally a data system, where generating and storing more data is essential to powering it—a dynamic that directly drives Western Digital's business. Despite recent market volatility around AI infrastructure stocks raising concerns, Tan argued that this demand remains durable, projecting a 25% compound annual growth rate (CAGR) over the next five years.

During the interview, Tan further broke down the four key drivers of storage demand he is watching: cloud growth, AI model training, inferencing and agentic AI systems, and the emerging opportunity in physical AI, which includes autonomous vehicles and robotics. These areas collectively form the basis for sustained demand for high-capacity storage devices.

Beyond the macro-strategic perspective, Tan also responded to specific operational questions, including customer selectivity, the company's competitive positioning, Western Digital's technology advantage in aerial density, and why vertical integration gives the company a durable edge in the market. For readers interested in the resilience of the AI hardware supply chain, this interview offers specific context for understanding the role of the storage industry in the AI cycle from an executive perspective.

According to reports by Techmeme and AI Weekly citing The Information, tenants of AI data centers—such as hyperscalers and AI labs—are now demanding late-delivery penalties exceeding $5 million per day, a significant increase from approximately $1 million in the spring of 2026.

This shift in risk allocation has prompted insurance broker Marsh to develop custom "delay-in-delivery" policies through its global digital infrastructure practice. Marsh's Nimbus facility raised its limits to $2.7 billion in January 2026. Goldman Sachs estimates that only about half of scheduled U.S. data center capacity will arrive on time.

Contracts such as OpenAI's deal with SB Energy illustrate this leverage: OpenAI receives one free day of rent for every day of delay, with penalties doubling after 90 days. This indicates that major AI buyers are using their bargaining power to transfer supply chain delay costs to developers.

South Korean battery maker LG Energy Solution announced on October 27 that it has signed a strategic business cooperation memorandum of understanding (Non-binding MOU) with global renewable energy company Enerig to advance energy storage system (ESS) projects in Europe and North America. Under the agreement, the two companies plan to achieve approximately 2GWh in supply targets within 24 months, with about 1.5GWh for the European market and 0.5GWh for the North American market. Additionally, LG Energy Solution stated it will explore expanding the total supply scale to 4.5GWh through its European and U.S. corporate entities.

Enerig is an Israel-headquartered independent power producer (IPP) involved in the development and operation of solar, wind, and energy storage projects, currently aiming to build a global asset portfolio exceeding 8GW. The first specific project under this collaboration will be the Jonava hybrid renewable energy project in Lithuania, which includes approximately 320MWh of energy storage facilities and is expected to begin commercial operations next year. Previously, LG Energy Solution supplied products to Enerig's Nové Čarnovo ESS project in Poland in 2024, establishing an initial cooperative relationship.

LG Energy Solution emphasized that this cooperation is significant for securing large-scale ESS project opportunities in Europe. Recently, the company has achieved notable progress in the North American region through AI data center, grid, and renewable energy-linked ESS businesses. By partnering with Enerig, LG Energy Solution aims to expand its business footprint from North America to Europe, thereby strengthening its position in the global energy storage market. Kang Chang-beom, Head of the ESS Battery Business Division at LG Energy Solution, stated that the company will leverage differentiated technology and local production capabilities to accelerate the growth of its global ESS business.

Naver Map, the leading mapping application in South Korea operated by tech giant Naver, announced on October 12 that it has officially enabled voice navigation support for Meta's AI smart glasses. This marks the first integration of a domestic map service with consumer AI glasses in the country, initially prioritizing compatibility with Meta's Ray-Ban and Oakley models.

According to hands-on testing reported by technology outlet ZDNet Korea, users can pair their Ray-Ban Meta glasses with the app once, after which they can set destinations in Naver Map and start navigation. The glasses then provide audio cues and turn-by-turn directions via built-in speakers, including distance to straight lines, turning points, and arrival notifications. The system was observed to quickly recalculate and guide users even when they deviated from the planned route.

This feature allows pedestrians to navigate without repeatedly checking their smartphone screens, freeing up both hands and attention. Beyond navigation, the glasses also allow users to capture photos and videos using a side button. Naver Map stated plans to expand this voice navigation service to other upcoming AI eyewear devices, such as Google and Samsung's 'Intelligence Eyewear,' and is exploring augmented reality (AR) visual path guidance for glasses equipped with displays.

Seoul AI Hub, a public institution supporting AI startups in South Korea's capital, announced plans to establish a new facility called the 'Baby Unicorn Center' next year. The center will specifically target AI companies that have secured cumulative investments of over 5 billion KRW (approximately $3.7 million USD), expanding support from early-stage incubation to post-Series B growth phases.

The center will offer scalable office space, priority allocation of GPU computing resources, and assistance with recruiting core talent. Additionally, Seoul AI Hub will partner with the Korea Angel Investment Association to facilitate investor relations (IR) events for overseas venture capital and connect companies with follow-on funding opportunities. Byeon Woo-seok, Director of Seoul AI Hub, stated that AI startups need simultaneous support in capital, talent, and infrastructure to sustain growth and enter global markets after passing the initial stage.

AI Engineer conference speaker Harish Gaggar, an engineer at Intuit Credit Karma, presents a detailed walkthrough of the infrastructure required to move enterprise AI agents from local development to production. The talk focuses on the 'paved path'—a standardized deployment process that ensures agents have consistent access to identity, memory, traces, evaluations, and secret handling.

Gaggar describes an architecture where a centralized Model Context Protocol (MCP) gateway sits between agents and tools, managing authentication and routing for every request. Custom MCP servers expose reviewed tools, while observability is maintained through OpenTelemetry, Tempo, Prometheus, and Grafana, allowing teams to monitor failures, latency, and access behavior.

The session details a LangGraph workflow that incorporates policy checks and human approval for sensitive actions. State management is handled through threads, runs, and checkpoints, with separate tracking tables recording model usage, tool calls, costs, and evaluation fields. Gaggar emphasizes that these records allow for inspection of agent behavior rather than establishing a validated quality score for the entire system.

Security and efficiency measures include a PII filtering layer designed to intercept sensitive inputs before they reach models or traces, though detection remains a control to be evaluated rather than a guarantee. Additionally, context optimization techniques such as pruning, reranking, deduplication, and compression are demonstrated, showing significant token reductions in specific queries, although no paired answer quality study or measured business ROI was presented.