Agents Need Context: Introducing Canvas Connectors, Fleet-wide AI Agent Visibility, and More

Honeycomb launched Canvas Connectors in general availability, offering 12 connectors including GitHub, PagerDuty, Jira and Amplitude that let the Canvas AI agent read from and act in systems outside Honeycomb's telemetry. Honeycomb also introduced AI Ecosystem in early access for fleet-wide visibility into AI agent performance and cost, made Anomaly Detection generally available, added onboarding via Honeycomb MCP from Claude Code, Cursor or Codex, and donated its adaptive tail sampling algorithms behind Refinery to the OpenTelemetry Collector.

Agents Need Context: Introducing Canvas Connectors, Fleet-wide AI Agent Visibility, and More Canvas agents can now read your code, incidents, runbooks and tickets, to get to the right solution, first time. Also new: AI Ecosystem and LLM cost tracking in early access, Anomaly Detection GA, and onboarding from your coding agent. By: Dan Juengst https://www.honeycomb.io/author/dan-juengst Bringing the Most Advanced Sampling to the OpenTelemetry Collector Honeycomb is donating its adaptive tail sampling processor, built on years of Refinery experience, to the OpenTelemetry Collector. See how adaptive sampling, trace fingerprinting, and sample rate attribution work, and how to try it today with the Honeycomb Collector Distribution. Read Now https://www.honeycomb.io/blog/bringing-most-advanced-sampling-opentelemetry-collector Earlier this year, we introduced more capabilities to support agents in production, a more chaotic, complex environment that requires a tremendous amount of context to understand https://www.honeycomb.io/blog/scary-things-in-production-context-helps-find . Unlike tools that capture shallow, pre-aggregated metrics, or cannot join a metric, trace, and log in one query, Honeycomb retains the context and connective tissue from telemetry data to build a nuanced view of production. AI-driven development has changed the nature of production incidents. There are more of them, they arrive faster, and they're stranger. Failures that don't throw an error, regressions with no obvious deploy to roll back, novel failure modes that surface in production first. Meanwhile, the teams we work with now run fleets of AI agents in production https://www.honeycomb.io/blog/introducing-ai-ecosystem . Honeycomb is rolling out new changes to adapt to the ever-increasing scale and complexity of production so you don't have to sacrifice reliability for innovation. Schedule a demo See the power of Honeycomb Intelligence. Speak with one of our experts today. What's new - Canvas enhancements: The Canvas agent is already your AI co-pilot to debug and investigate the complexity of production, collaborating and surfacing findings in a multi-player workspace for humans and agents. Here's how it's getting better: - Canvas Connectors GA : 12 connectors and growing, from GitHub and PagerDuty to Jira and Amplitude, that let the agent in Canvas read from and act in the rest of your stack, because there's more context beyond telemetry. - Canvas extensions: edit triggers, SLOs, and boards from Canvas, and draw on charts to scope the agent's work. - AI Ecosystem early access : a fleet-wide view of AI agent performance and cost. - Anomaly Detection GA : learns each service's normal error rate and presence, with nothing to configure manually. - Onboard with Honeycomb MCP: connect your codebase from Claude Code, Cursor or Codex and instrument with a guided prompt. - Adaptive tail sampling in OpenTelemetry: we're donating the dynamic sampling algorithms behind Refinery to the OpenTelemetry Collector https://www.honeycomb.io/blog/bringing-most-advanced-sampling-opentelemetry-collector , so teams keep the important context without blowing through their budgets. Canvas Connectors: putting hi-definition telemetry in full context Back in May, we launched the new Canvas https://www.honeycomb.io/blog/honeycomb-canvas-multiplayer-workspace-for-agentic-era , a live workspace where humans and agents investigate production together. Today, we're shipping the next big piece. Connectors are how Canvas reaches context outside Honeycomb's own telemetry: your GitHub repos, PagerDuty incidents, Confluence runbooks, or a Jira ticket someone filed. Canvas isn't your typical “co-pilot” agent. It's designed with the realities investigating complex production environments. Sure, there's natural language chat, but it responds with evidence, including visuals like BubbleUp https://www.honeycomb.io/platform/bubbleup graphs. It's a multiplayer experience that keeps teams on the same page as new experts join an investigation, and it comes with expert observability skills https://docs.honeycomb.io/investigate/canvas/skills built-in-skills . “Canvas gives you that unified look of ‘This is the problem I'm trying to solve, these are the queries I threw at it, the questions I gave it.’ You can see the thought process. You see the charts that it's plotting out.” George Luong, Senior Engineering Manager at Slack In other words, Canvas is how to paint a collaborative picture of what's going on in production. Today, we're adding more colors to Canvas' palette. Why telemetry is never the whole picture Wide event telemetry gets you a long way. Honeycomb can tell you the exact moment latency started to increase, which endpoint, which customers, which spans, and that it lines up with a deploy marker. That's the when and the where . But telemetry cannot tell you what was in the deploy, that someone is already fifteen minutes into an incident on the upstream service, or that your team wrote a runbook for exactly this failure after the last three times it happened. That context lives in GitHub, PagerDuty, Confluence and Jira. Today, an on-call engineer gathers it by hand, across six tabs, at 2 a.m. An agent that only sees telemetry will propose the telemetry-shaped fix: revert the deploy, restart the service, kill the query. Sometimes that's a good enough solution, but it's not always the best one. What a connector does Every connector declares what it's allowed to do, in one or more of three jobs. Evidence in. The agent can search, read, and cite the tool during an investigation. A PR diff, an incident timeline, a runbook step: each lands on the canvas as a card, labeled with the connector it came from, next to the chart or trace it explains. Work out. The agent can draft or take an action in the tool: file a ticket, update an incident, post a reply. Reads are allowed by default once you connect a tool. Writes ask for permission each time. The agent does the legwork, and a human approves the action. Impact context. The agent can translate a finding into a business number from real data instead of a rough estimate: revenue at risk, the value of affected accounts, which user segments are actually feeling it. For example, with Amplitude connected, that number can come from your product analytics. Connectors are built on MCP, the same open protocol behind Honeycomb's own MCP server https://www.honeycomb.io/technologies/ai-agents , which is why the roster below will keep growing. Canvas Connectors in action A FinTech payments team ships a new build. p99 latency on transfers jumps 12x and the SLO starts burning. Honeycomb triggers an auto-investigation, lines the regression up with the deploy marker, and shows each transfer now making nine sequential calls to a screening service instead of one. Based on the telemetry alone, the agent can only suggest reverting the build. With GitHub connected, the agent sees the build is actually three merged PRs. One introduced the per-payee loop that's causing the latency issue. Another is a compliance fix that a regulator expects to go live today. Without that extra context, anyone might have chosen to roll back the whole build and put the compliance fix at risk. Instead, Canvas can recommend a six line hotfix in one PR to solve the issue without reverting other necessary updates. The production telemetry was right, but extra context from the Canvas Connectors led to a better, more targeted response. Also new in Canvas You can now create or update triggers, SLOs, and boards from Canvas https://changelog.honeycomb.io/edit-triggers-slos-and-boards-in-canvas-343209 . Ask for an SLO on the endpoint you just investigated, review the definition, and approve it. Edits always require human approval and aren't available to the agent during automatic investigations. Multi-agent observability: zoom in on one conversation from the fleet-wide view One of the things customers love about Honeycomb is the ability to go from a high-level view, like a p99, and drill all the way down into specific sessions outside that band. Most tools show you that high-level metric but leave you in a dead end wading through logs to look for the details. Honeycomb is different because it builds up that higher level aggregate from wide events that retain the full context of an event. The same applies to how Honeycomb builds a view of what your agents are doing in production. Agent Timeline https://www.honeycomb.io/blog/agent-timeline-generally-available answers “What happened inside this AI agent's conversation?” It traces every agent invocation, LLM call, tool call, and downstream trace, all bound by a conversation ID. Today, we're adding AI Ecosystem, in early access, which answers “What's happening across all my agents, and where do I look first?” AI Ecosystem performance view AI Ecosystem gives you a continuous, aggregate pulse on failure rates, retry trends, latency, token usage, and cost across every agent you run. Which agents are trending worse this week? Is that retry spike one agent or the whole fleet? We pushed a bad update, how many conversations were affected? Today, the answer to each of those is a person writing a bespoke query. AI Ecosystem makes it a five-minute review, with one click from any number to the conversation on the Agent Timeline that explains it. It runs on the same span telemetry https://www.honeycomb.io/blog/instrumenting-ai-agents-agent-timeline-opentelemetry-guide Agent Timeline already uses, so there's no new instrumentation and no new setup. AI Ecosystem cost view The cost view in AI Ecosystem gives you average cost per conversation as a baseline, then breaks total and average spend down by agent, model, agent and model together, and token type. Any aggregate is one click from the conversations behind it, and you can ask Canvas about cost trends directly. These are estimates derived from the model and token counts on your spans and public list prices: use them to see what's driving spend, not to reconcile your provider invoice. Learn more about AI Ecosystem https://www.honeycomb.io/blog/introducing-ai-ecosystem . Anomaly Detection is now generally available Anomaly Detection https://docs.honeycomb.io/notify/anomaly-detection/get-started get-started-with-anomaly-detection learns a statistical baseline for each of your services and surfaces deviations as they happen. There's nothing to configure: Honeycomb identifies which services have enough steady traffic to model reliably, onboards them automatically, and shows you which are covered and which aren't. When an anomaly begins, you're notified in Slack or through your configured recipients, and an auto-investigation kicks off in Canvas so the first stretch of the work is done before you've even logged on. Request rate, latency, and seasonality-aware detection are coming soon. Onboard with Honeycomb MCP Getting data into Honeycomb now starts wherever you already work with AI. Connect your codebase through the Honeycomb MCP server https://changelog.honeycomb.io/new-onboard-with-honeycomb-mcp-343984 from Claude Code, Claude Desktop, Codex, Cursor and others, and a guided prompt walks you through auto-instrumentation step by step. Send live data from your app, generate synthetic data from your code to try things out, or explore data that's already there. Honeycomb then generates a Canvas investigation so you can start asking questions. Go to Manage Data Send Data to get started. Adaptive tail sampling in OpenTelemetry Sampling is a standard practice for managing event volume in an observability pipeline, and different methods carry different tradeoffs. As part of this launch, we're donating adaptive tail sampling to the OpenTelemetry Collector: the dynamic sampling algorithms that Refinery https://www.honeycomb.io/platform/refinery-as-a-service , Honeycomb's open source sampling proxy, is built on. Unlike the more rigid rules in the Collector's current tail sampler, it adjusts dynamically, so teams keep the important context without blowing through their budgets, even during spikes. Refinery is still the more complete solution, adding clustering, scaling, and routing on top, but bringing its core sampling approach upstream helps the whole industry adopt these capabilities. Read more and learn how to get started with the current distribution https://www.honeycomb.io/blog/bringing-most-advanced-sampling-opentelemetry-collector . An increasingly multi-agent world Agent fleets are growing, and so is the number of internal agents in use. It's clear that the future isn't built on building “one agent to rule them all.” Teams of agents and humans will need to collaborate with shared context. That requires letting agents like Canvas pull in more context to make more nuanced assessments, and it requires retaining the context between an aggregate view of agent behavior and individual agent conversations. This release expands shared context in several directions, building on Honeycomb's architecture to store the full context of every event. Canvas Connectors enrich that context to propose the right fix the first time, and, with permission, to carry it out. AI Ecosystem builds an aggregated view without dropping the details so teams can dynamically explore an agent's health in the context of the fleet. See the latest from Honeycomb in action at O11yDay London on October 6th and 7th https://www.honeycomb.io/events/o11yday-london-2026 . Get started A dozen Canvas Connectors are already available to all Honeycomb customers today under Settings Connectors ; the documentation https://docs.honeycomb.io/investigate/canvas walks you through connecting your first tool. AI Ecosystem and LLM cost tracking are in early access. Anomaly Detection is already available for every team. If you're not on Honeycomb yet, we'd love to show you what an investigation looks like when the agents and your team are working from shared, deep context and expert skills. Schedule a demo See the power of Honeycomb Intelligence. Speak with one of our experts today.