January 2026 Summaries

Nebulock is revolutionizing threat hunting by using AI agents that actively search for threats within an organization's environment, aiming to democratize the process beyond traditional rule-based systems. These agents, which can operate autonomously or alongside analysts through natural language interaction, are designed to identify and address threats with clear outcomes. Nebulock places a significant emphasis on transparency, showcasing the internal reasoning of its agents to customers, aided by the Arize platform for evaluation and confidence in agent outputs. The company collaborates with expert threat hunters to integrate their extensive knowledge into the AI agents, ensuring a robust development process. A key lesson learned is the importance of a strong experimentation and development flow to enable rapid iteration and improvement. Nebulock uses Arize AX for comprehensive observability and evaluation, benefiting from its ability to track experiments and monitor production performance, which aids in building reliable and efficient threat-hunting solutions.

The exploration of hierarchical memory management in AI agents highlights the effectiveness of using Unix-based file systems as a memory layer, drawing parallels to historical computing memory solutions. This approach leverages simple Unix commands like grep, ls, and glob to dynamically manage and expand the context window for agents, allowing them to handle large datasets efficiently. The use of these composable tools, as demonstrated in the productivity tools Cursor and Claude Code, enables agents to perform complex tasks by creating ad-hoc query plans and dynamic indexes on-demand. This method contrasts with traditional database indexing by offering flexibility and the ability to handle diverse patterns without pre-built structures. The insights gleaned from building the Alyx agent at Arize AX underscore the potential for integrating database access with file system operations for managing vast data volumes, suggesting a future where agent memory systems are built on a hierarchy of composable, Unix-inspired tools.

The OWASP Top 10 for Agentic Applications: Compliance Guide, published in January 2026, outlines key security risks and mitigation strategies for agentic AI systems, focusing on observability features and metrics provided by Arize AX. It aligns the OWASP Agentic Security Initiative's top ten risks, such as agent goal hijacking, tool misuse, identity abuse, and supply chain vulnerabilities, with monitoring and evaluation techniques to detect and address these threats. The guide emphasizes the importance of tracing, monitoring, and setting up evaluators to safeguard against risks like unexpected code execution, memory poisoning, insecure inter-agent communication, cascading failures, human-agent trust exploitation, and rogue agents. It highlights the necessity of version control, approval workflows, dashboards, and alert systems to maintain the integrity and security of AI systems, while also providing practical steps such as tracing agent actions, evaluating tool usage, and monitoring communication and execution patterns.

As AI agents are increasingly deployed in production environments, they encounter challenges due to conditions not covered during their training, leading to operational failures characterized by recurring patterns. These failures include retrieval noise, hallucinated arguments in tool calls, recursive loops, and guardrail failures, among others. Variability introduced by AI agents contrasts with the repeatability expected in traditional software, posing risks such as agents fabricating responses or making inefficient decisions that inflate operational costs. Misalignment between pre-trained biases and contextual information can result in inappropriate responses, which is exacerbated by unhandled external API changes and instruction drift in long sessions. AI agents' non-deterministic nature necessitates robust monitoring and guardrails to intercept potentially harmful actions, and tools like Arize AX are suggested for mapping decision paths and ensuring functional safety through trajectory evaluations.

Observability-driven sandboxing is a method for securing AI agents by enforcing runtime policies on their actions, ensuring safety without altering the agent's planning capabilities. This approach uses Google ADK and Arize Phoenix to intercept tool calls, treating them as capability requests that are evaluated against explicit policies at execution time, with decisions being traced using OpenTelemetry. The sandbox acts as a mediator, allowing or blocking actions based on predefined rules, such as restricting file access to designated directories or limiting network connections to approved hosts. Every decision made is recorded and visualized, offering a transparent, auditable execution trace that helps developers understand why certain actions were allowed or denied. This framework not only enhances security by preventing unauthorized actions like file modifications or network breaches but also maintains the agent's reasoning abilities. By adopting such observability-first approaches, developers can manage AI agents more reliably in complex environments, ensuring both safety and control without compromising on transparency or auditability.

In 2026, the debate over the most effective interface for AI agents—whether filesystem, API, or database—remains unresolved, reflecting the experimental nature of current agent architecture. A benchmark by Letta revealed that a filesystem-based agent excelled in memory tasks, sparking discussion among experts at Arize about the inherent limitations of APIs and the potential of agent-to-agent communication. Tony Powell and Laurie Voss highlighted that while filesystems perform well due to pretraining familiarity, they are not inherently superior; rather, they expose the shortcomings of API interfaces for AI. Aman Khan reframed the debate by distinguishing between interface and deployment models, noting that options like filesystem interfaces with remote data and runtime materialization, as used by Phoenix Insight, offer a promising hybrid approach. This method allows agents to utilize familiar filesystem tools without the burden of local data maintenance, though it acknowledges that the field is still in an exploratory phase, with future developments likely to refine these models further.

Integrating the Arize AX MCP server with Google's Antigravity allows developers to instrument AI applications with tracing and observability without writing any code. This integration introduces a streamlined workflow that enables debugging, monitoring, and improving AI agents directly within the development environment. By adding the Arize AX MCP server, developers can leverage OpenTelemetry-based tracing with curated best practices and examples. The process is facilitated by Antigravity's ability to analyze code, identify necessary dependencies, and resolve any conflicts that arise, such as version mismatches. The system generates artifacts throughout the process, providing transparency and documentation for every step, ultimately allowing the developer to verify and implement changes confidently. This approach transforms manual AI observability into an autonomous workflow, enhancing efficiency and clarity in managing AI applications.

The essay by Jaya Gupta and Ashu Garg highlights the emerging significance of capturing decision traces and transforming them into queryable context graphs, suggesting this will be the next major data advantage for enterprises. This approach emphasizes understanding the rationale behind actions, not just documenting actions themselves, which resonates with observed trends where agents integrate structured systems with human-generated context to inform dynamic business decisions. A notable example is Cursor, an agent-centric pattern in software development, which synthesizes data across diverse systems to generate actionable insights and diagnostics. This trend is extending beyond software development into areas like DevOps and SecOps, where agents utilize both traditional systems and underutilized data sources such as Slack threads and emails to form and execute decisions, effectively shifting the locus of decision-making. As these agent-based systems evolve, organizations are treating agent traces as durable business assets, integrating them into data lakes for analysis and feedback, which could potentially redefine the landscape of systems of record by prioritizing business reasoning as a core asset. The future holds questions about control and interfaces for this reasoning data, with implications for who captures value and what systems of record will look like.

Arize AX has introduced several updates to enhance its platform, including multi-span filters and improved playground views. The multi-span filters allow users to apply complex conditions using logical operators and relationship filters to analyze traces more deeply. The improved playground views, now called Playgrounds, enable users to save and share configurations of their sessions, preserving work and facilitating collaboration among teams. Additionally, new spaces created from December 5, 2025, will automatically use real-time ingestion, ensuring immediate visibility into production workloads. The platform also expanded its library with new cookbooks for tasks such as tracing AI agents and evaluating AI at scale, and increased model support to include GPT-5.2, Opus, and Haiku 4.5 models for both playground and online tasks.