September 2026 Summaries

Honeycomb’s design enablement team refreshed the product UI to address years of accumulated visual inconsistency and complexity caused by incremental feature additions and one-off design decisions. Updates to the Lattice design system include refined shadows for clearer hierarchy, rounder controls to reduce visual severity, and motion tokens that make interactions more consistent and communicate system activity. The refresh also exposed places where teams had bypassed shared components and tokens, creating an opportunity to reduce design debt through broader migrations. Heatmaps received a new magma-inspired, brand-aligned color ramp that improves distinction between values, supports dark mode, and performs better with colorblind filters, while feedback continues to inform contrast refinements. Canvas was moved higher in the sidebar to make investigations easier to start, and new sizing, spacing, and layout tokens support a denser but more organized interface.

Honeycomb announced updates aimed at helping teams manage increasingly complex production environments and fleets of AI agents by preserving detailed telemetry context and connecting it with external operational tools. Canvas Connectors, now generally available, integrate services such as GitHub, PagerDuty, Jira, Confluence, and Amplitude so Canvas can incorporate evidence, propose targeted fixes, estimate business impact, and draft actions while requiring human approval for writes. The company also introduced AI Ecosystem in early access, providing fleet-wide views of agent performance, latency, retries, token usage, and estimated costs, with links to detailed agent conversations through Agent Timeline. Other releases include generally available automatic Anomaly Detection, guided instrumentation through the Honeycomb MCP server for AI coding tools, Canvas-based editing of triggers, SLOs, and boards with approval controls, and a contribution of adaptive tail-sampling algorithms to the OpenTelemetry Collector. These capabilities are intended to give human teams and agents shared context for faster, more precise investigations without losing detail or exceeding observability budgets.

Honeycomb announced AI Ecosystem, an early-access observability capability for monitoring fleets of production AI agents across performance, failure rates, token use, latency, costs, and individual conversations. Built on Honeycomb’s high-cardinality data model, the feature is intended to help engineering teams identify degraded agents, distinguish regressions from normal variation, assess customer impact, and investigate specific problematic interactions; general availability is expected by the end of 2026 or early 2027. The company also expanded its collaborative Canvas investigation workspace with MCP connectors for tools including GitHub, Linear, Pylon, incident.io, Atlassian, Braintrust, Sentry, Amplitude, and Checkly, allowing teams to combine telemetry with code, tickets, design context, and team discussions during incident response.

Honeycomb has introduced AI Ecosystem in early access to extend its AI observability platform from individual agent conversations to fleet-wide monitoring. Building on Agent Timeline, which traces LLM calls, tool invocations, handoffs, and downstream spans within a single workflow, AI Ecosystem helps teams operating many production agents identify performance regressions, failure patterns, deployment blast radius, and cost drivers across their entire fleet. Its Agent Fleet Performance capability summarizes health indicators such as failure rates, retries, latency, and volume, while LLM Cost Tracking provides estimated spending by agent, model, token type, conversation, and individual LLM call using public pricing data. Users can drill from aggregate metrics directly into the relevant Agent Timeline conversation without rebuilding queries or switching tools, using existing span telemetry and Honeycomb’s high-cardinality, context-rich data model rather than new instrumentation or pre-aggregated dashboards.

Honeycomb’s adaptive_tail_sampling processor brings adaptive trace sampling algorithms from Refinery and dynsampler-go to the OpenTelemetry Collector, aiming to retain important traces such as errors and rare routes while preserving statistically accurate traffic estimates. Unlike fixed head sampling, it buffers traces until completion signals or limits are reached, applies ordered OTTL-based rules, and records each decision as a W3C TraceState sampling threshold so downstream tools can reweight sampled data correctly. It supports adaptive percentage sampling for a target share of traffic, adaptive throughput sampling for a span-volume budget, and fixed or always-keep rules, with fingerprints such as service name and HTTP route used to distinguish traffic classes. The processor handles late spans through a decision cache, makes rule-based decisions during buffer evictions, drains buffered traces on shutdown, and can compose with existing OpenTelemetry samplers so the strictest effective probability is retained. Benchmarking indicated high throughput and bounded memory under overload, though deployments require routing all spans for a trace to the same Collector instance, rules require restarts to change, and memory limits are based on trace and span counts rather than bytes. Available in beta in Honeycomb’s Collector distribution and progressing upstream, the processor offers Collector-native adaptive sampling, while Refinery remains positioned for more mature operational requirements such as clustered scaling, live rule changes, overload protection, and multi-tenant operations.

Honeycomb presents Canvas as a collaborative, AI-assisted workspace intended to help teams establish a continuous improvement loop for production AI agents by instrumenting behavior, investigating runs, identifying systemic failures, implementing changes, and validating results. The approach relies on OpenTelemetry GenAI conventions to capture traces, conversations, agent identities, operations, token usage, prompts, tool activity, errors, and evaluation scores, enabling Honeycomb’s Agent Timeline and broader cost, latency, and quality analysis. Canvas can examine individual conversations or patterns across thousands of runs, prioritize issues by severity and frequency, and use integrations with GitHub, Linear, and Slack to connect findings to code, tickets, pull requests, and deployment history, with approval required for write actions. It also supports before-and-after version comparisons, dashboards, triggers, automated investigations, custom skills, evaluation datasets, and live collaboration to detect regressions and preserve investigation context. Honeycomb says it uses Canvas to improve Canvas itself, while noting that connectors such as GitHub and Linear are in beta and expected to reach general availability in fall 2026.

Honeycomb’s concluding statement on AI norms presents artificial intelligence as a powerful but subordinate tool whose use should strengthen human judgment, accountability, communication, and outcomes rather than replace them. It argues that employees remain responsible for work produced with AI, should distinguish between quick assistance and rigorous collaboration, and must respect colleagues’ time by reviewing generated material, making specific requests, and explaining an artifact’s purpose and expected feedback. The document differentiates personal and professional opinions, collaborative written artifacts, and software artifacts, emphasizing that AI use is least appropriate when communication depends on an individual’s authentic voice, such as personal messages, feedback, or performance reviews. It encourages teams to resolve AI-related frustration through explicit discussion of expectations, effort, provenance, and consent instead of rigid rules, while assigning managers particular responsibility for preserving trust and exercising independent judgment. Honeycomb also acknowledges AI’s environmental, intellectual-property, privacy, bias, and labor-related harms, commits to ethical vendor review, attribution and compensation where appropriate, bias awareness, fair pay, training, and nonretaliation for lawful employee activism. It concludes that careful AI adoption can help the company improve products and compete effectively while maintaining humane workplaces, strong customer relationships, and mutually beneficial outcomes.

As organizations deploy agentic AI systems, they face rising observability costs because traditional metrics, logs, and traces are often collected, stored, and queried separately despite overlapping information. The text argues that AI agents require especially rich context—including models, prompts, tool calls, failovers, conversation turns, and response timing—because their nondeterministic behavior makes troubleshooting and long-term pattern analysis more difficult. It presents “wide events,” or arbitrarily broad structured records that can include trace identifiers and other contextual fields, as an alternative single data format from which teams can derive metrics, logs, traces, service-level objectives, and agent timelines at query time. By retaining raw, contextual events in one datastore, proponents contend that organizations can reduce duplicated ingestion and computation costs while avoiding a tradeoff between context and preaggregation, although conventional metrics may be cheaper in limited predictable cases. The piece recommends examining duplicated telemetry across observability systems, determining which data must remain at request level, and relating AI token and model activity to telemetry growth to better forecast spending, while promoting Honeycomb as a platform designed around the wide-event approach.

Honeycomb relational queries enable teams to investigate errors across an entire distributed trace rather than relying only on attributes attached to the initially selected event. Starting with checkout-service spans where `error = true`, the example progressively combines child exception messages, root-level URLs and HTTP status codes, and user, shipping, and payment information located on separate spans to create a single business-focused error view. Relational prefixes such as `root`, `parent`, and `child` target directly related spans, while `any`, `any2`, and `any3` retrieve attributes from up to three different matching spans elsewhere in the trace; `any` fields require corresponding WHERE conditions and each prefix must match attributes on the same qualifying span. The `none` prefix excludes traces containing specified attributes, and grouping by trace IDs can make results easier to inspect. These queries can be used in Honeycomb boards, triggers, SLOs, agents, and shared query links, helping teams correlate technical failures with customer and financial impact.

Honeycomb’s engineering leadership reflects on its year-long AI adoption mandate, arguing that AI is essential for shaping industry practices, maintaining credibility as an observability expert, and helping customers move beyond outdated software-development approaches. The organization’s north star is to become among the top 10% most AI-enabled and productive engineering teams at comparable companies, alongside a company-wide goal of doubling impact through AI. Near-term objectives include safely auto-merging at least 25% of AI-reviewed pull requests while keeping change failure rates below 3%, maintaining service-level objectives and sustainable incident workloads, adapting work to remain meaningful and sustainable, and publicly sharing lessons learned. Honeycomb emphasizes that AI should support the entire software development lifecycle rather than merely accelerate coding, with employees expected to experiment, develop shareable insights, and work with managers to assess performance rather than meet fixed individual usage targets. The guidance supports exploration of agentic workflows and possible software-factory models without prescribing exact practices, while encouraging engineers to use discussions productively, avoid unproductive Slack debates, and acknowledge the professional and emotional adjustments that accompany more autonomous AI tools.

Honeycomb is donating an adaptive tail sampling processor to the OpenTelemetry Collector, aiming to provide a more flexible alternative to traditional head, probabilistic, and static tail sampling approaches. Unlike conventional tail samplers, the processor buffers spans to evaluate full traces while making decisions soon after receiving a root span, helping reduce memory and CPU use. It combines configurable static rules, such as always retaining errors or dropping health checks, with trace fingerprinting that distinguishes traffic by attributes such as services, routes, response codes, and tenant identifiers. Adaptive rates are recalculated periodically to distribute either a percentage-based or throughput-based sampling budget across fingerprints, preserving visibility into low-volume requests while reducing data from high-volume paths and limiting costs during spikes or attacks. The processor also records sampling thresholds in OpenTelemetry trace state to support accurate backend extrapolation, and it is available ahead of upstream release through the Honeycomb OpenTelemetry Collector Distribution for deployments using OTLP, Kubernetes, or Docker.