March 2026 Summaries

Honeycomb has expanded its hosted MCP server to integrate deeply into users' development and observability workflows, allowing them to build agentic workflows across various tools and adapt them to their preferred agents like Claude Code or Cursor. To enhance these workflows, Honeycomb has released open-source Agent Skills that distill their expertise in observability into eight skills and two autonomous agents, which help users get started with observability, migrate to OpenTelemetry, or troubleshoot production issues. These skills guide users through the full journey of setting up instrumentation, performing gap analyses, and conducting structured production investigations using tools like BubbleUp and SLOs. An Instrumentation Advisor agent is included to bridge users' codebases with Honeycomb data, providing guidance on what to instrument and how to maintain distributed traces during migrations. The release aims to equip developers with not only the APIs but also the knowledge needed to effectively use them, signifying a shift in the landscape of developer tools toward more knowledgeable and autonomous agents.

In large-scale production systems, anomalies such as unusual request patterns and traffic spikes are commonplace, appearing as noise without detailed analysis. While AI and machine learning can detect anomalies, identifying which ones require action is more challenging, often necessitating an understanding of developer intent and the nuanced context of each system change. This involves tracking changes in production environments, where every deployment is unique, and leveraging high cardinality data to trace the effects of these changes over time. The value of context is emphasized, as it exponentially increases the power of data analysis by allowing for the combination of multiple attributes, revealing insights that were previously obscured. A case study of Homeaglow illustrates how detailed instrumentation and telemetry can surface hidden issues and inform product decisions, demonstrating the importance of treating data collection and analysis as a product problem rather than merely an infrastructure one. The future of this approach includes leveraging agents that require even more context and cardinality, as the tech industry transitions towards automated, factory-like software development processes.

Engineering leaders often equate identity diversity with cognitive diversity, assuming varied backgrounds lead to different perspectives, yet this assumption can be misleading in understanding complex systems. Research indicates that cognitive diversity, or the variety of mental models individuals bring to problem-solving, is more crucial than identity diversity, especially in navigating the intricacies of sociotechnical systems like production systems, which are characterized by unpredictable behaviors and complex interdependencies. To harness cognitive diversity, organizations should intentionally compose teams with varied mental models and encourage independent analysis before group discussions to capture diverse perspectives. This approach is particularly vital as AI integration increases system complexity, where traditional mental models may not suffice. Emphasizing cognitive diversity helps stress-test solutions and avoid the pitfalls of relying on homogeneous perspectives, ultimately fostering more accurate and effective problem-solving in engineering environments.

In early February, a small, invite-only unconference was held in Deer Valley, Utah, led by Martin Fowler and Thoughtworks, to explore the evolving landscape of software engineering in the AI-native era. The event's key insights were summarized into ten themes, highlighting the emphasis on relocating rigor in software engineering from traditional constraints to more upstream areas like specification reviews, test suites, type systems, and risk mapping. However, the summary raises concerns about the exclusion of production systems from this rigor relocation, arguing that production is crucial as it represents reality where users interact with the software. The piece critiques the industry's persistent oversight of production environments, which are often seen merely as a site for bug fixes, and stresses the importance of observability and continuous monitoring in understanding software performance and user interaction. With AI accelerating the rate of change and reducing the cost of migrations, the need for robust real-time feedback loops and rich data ecosystems directly linking development and production has become more critical, suggesting a paradigm shift towards integrating production observability as a core component of software development practices.

In the context of software development, the text discusses the strategic decision of when to generate metrics within the Software Development Life Cycle (SDLC), suggesting a shift to the right, contrary to the typical shift-left approach. This shift involves generating metrics later in the telemetry pipeline, such as in the OpenTelemetry Collector, rather than within the application itself. The choice of when and where to generate these metrics is dependent on various factors, including the type of data, system size, and application type, with a focus on separating the measurement of data from the generation of metrics. This approach aims to improve operational efficiency, reduce costs, and enhance the granularity and accuracy of data, even if it sacrifices some real-time accuracy. The text highlights the use of tools like the signaltometrics connector to convert logs and traces into metrics more efficiently, emphasizing the benefits of exponential histograms for greater data granularity. Ultimately, it encourages considering different methodologies for generating metrics to optimize feedback and resource use in the development process.

Observability tools in the AI era are being evaluated differently due to their reliance on AI agents, which require fast, complete, and cost-effective data access to function effectively. Traditional evaluation criteria, such as dashboard quality and alerting flexibility, are being supplemented with new requirements like data completeness, query speed, and infrastructure integration. AI agents need high-cardinality data and real-time access to provide accurate and actionable insights, making the underlying data model and query infrastructure crucial factors in tool selection. Honeycomb is highlighted as an observability platform that aligns with these needs by offering a comprehensive data model, fast query processing, and robust integration capabilities, all without penalizing users for data richness. It's suggested that potential buyers focus on the practical aspects of data accessibility, query performance, and cost implications rather than being swayed by impressive but potentially misleading demo environments.

During a Black Friday event, a checkout latency issue revealed the limitations of traditional time series metrics platforms, which struggle with high cardinality issues and require pre-defined queries. This incident highlighted the need for a more flexible observability approach, especially as AI plays a greater role in software engineering. Honeycomb Metrics addresses these challenges by integrating time series storage with an events model, allowing for comprehensive data queries without the prohibitive costs associated with high-cardinality data. This dual-model system enables engineers to quickly diagnose and resolve complex issues by maintaining the necessary context in the data, ultimately improving the efficiency and effectiveness of both human and AI-driven investigations. As systems become more complex and less predictable, Honeycomb's approach allows for real-time, nuanced analysis, accommodating the evolving needs of modern engineering teams.

The Honeycomb Slackbot is a newly launched tool aimed at reducing the cognitive overhead engineers face when switching between multiple platforms during complex tasks by integrating observability features directly into Slack. Powered by Honeycomb Intelligence, this Slackbot allows engineers to ask questions and receive real-time data insights within Slack channels without needing to leave the platform, thereby maintaining focus and enhancing collaborative problem-solving. By tagging @Honeycomb Canvas, teams can transform Slack discussions into ongoing investigations, enabling them to access telemetry data and collaborate efficiently during incidents, onboarding, and everyday debugging. This integration not only streamlines incident response by providing quick access to relevant metrics and logs but also aids new engineers in understanding production systems through instant data-driven responses. Currently in beta for enterprise customers, the Slackbot aims to embed observability into the engineers' existing workflow, making it a valuable AI-powered teammate that supports seamless collaboration and data-informed decision-making.

Honeycomb.io has announced new AI-driven features and major product milestones, including the general availability of Honeycomb Metrics and expanded Model Context Protocol (MCP) integrations to accommodate AI development tools. As AI agents increasingly contribute to software production, Honeycomb aims to be the first observability platform tailored for AI by providing the necessary structured data and platform access for autonomous monitoring and optimization. New capabilities include Honeycomb Agent Skills for various AI agents, Automated Investigations that mimic skilled SRE responses, and a Slackbot that integrates AI-powered insights directly into Slack. Honeycomb Metrics offers a unified platform for both time series and event-based models, allowing teams to capture detailed metrics without incurring high costs associated with high-cardinality data. The expanded MCP capabilities facilitate embedding observability into AI tools, enhancing evidence-based investigations and fostering trust in AI-generated insights. With promotional pricing available until June 2026, Honeycomb continues to support engineering teams in managing growing telemetry volumes efficiently.

Honeycomb's MCP (Magic Customization Platform) allows users to quickly create tailored dashboards for their applications, services, or teams by leveraging data that their software sends to Honeycomb. Users can configure the platform via their IDE or command line, specifying detailed requirements like service health metrics, latency, error rates, and dependencies. The agent, using knowledge of the user's code and data, constructs a dashboard by examining code for service names, custom attributes, and relevant business metrics, and by analyzing existing data schemas and frequently-used queries. This results in a personalized dashboard that can be adjusted further after its creation. To enhance the process, users can download or edit skills that guide the agent in understanding user preferences better, ensuring the final dashboard aligns with their specific needs. Honeycomb also supports integration with AI agents like Claude Code and Cursor, requiring only a Honeycomb account with Intelligence enabled, and can work even if the software initially lacks data instrumentation, thanks to OpenTelemetry support.

A recent webinar on AI-assisted development with DORA delved into several complex and pressing questions from the audience regarding the challenges and opportunities presented by AI coding tools. Observability was a key theme, emphasizing that AI agents should be monitored like any other software using traces, logs, and metrics, and that the integration of AI does not introduce new types of complexity in resolving production incidents but rather increases the existing complexity and rate of issues. The discussion also highlighted the importance of maintaining strong engineering practices such as robust testing, documentation, and feedback loops to manage the rapid code generation by AI, while also suggesting that AI can speed up code reviews by handling routine checks before human review. For integrating AI into existing codebases, starting with new projects or improving existing documentation and test coverage were recommended strategies. AI's role in onboarding junior engineers was also discussed, noting its potential as a supportive tool that allows juniors to ask questions and explore solutions independently before seeking human mentorship, thus identifying documentation gaps and improving team knowledge.