The short version: Honeycomb — the production-observability company built on high-cardinality event data — pointed its platform at AI agents in May 2026. It reads the OpenTelemetry GenAI semantic conventions (v1.40.0), so there's no proprietary SDK and no framework lock-in, and its Agent Timeline renders a multi-agent, multi-trace run as a single view. If your agent already emits OTel spans, you can watch it in the same place you already watch your app.
What it is#
Most tools that arrived to "observe agents" started life as LLM-eval platforms — they grade a prompt and its output. Honeycomb comes from the other direction. It's an APM-lineage observability platform (co-founded and led by CEO Christine Yen), the kind of tool you already point at your database, your queue, and your request latency. In May 2026 it extended that same engine to treat agents as first-class citizens.
The distinction matters because an agent failure is rarely just a bad model turn. It's a tool call that timed out, an MCP server that returned a partial result, a retrieval query that hit a cold index — and the model reacting to all of it. You want the agent's reasoning and the system underneath it in one trace. That's the gap Honeycomb is aiming at.
Who it's for#
The solo founder or small team that is already emitting OpenTelemetry spans and doesn't want a second, bolted-on dashboard for the AI half of the stack. If you've resisted proprietary agent SDKs precisely because you don't want to re-instrument every time a spec moves, this is built for you: you instrument with plain OTel, and Honeycomb reads model evaluations, tool executions, MCP calls, LLMs, and agents automatically.
How to start#
- Instrument with OpenTelemetry. Emit GenAI spans from your agent — most frameworks and the OTel GenAI instrumentation libraries already do this. No Honeycomb-specific SDK required.
- Point your OTLP exporter at Honeycomb. Set the endpoint and your API key; spans start landing in seconds.
- Open Agent Timeline. It stitches a multi-agent, multi-trace workflow into one view — every LLM call, tool invocation, MCP call, and downstream system impact, in real time.
- Let Auto-investigations run. When an alert fires, Honeycomb gathers the relevant data and proposes a remediation, so triage doesn't start from a blank query.
The features that matter#
- Agent Timeline — the headline. One timeline for a whole multi-agent run, so a supervisor→worker handoff that stalls three hops deep is legible instead of scattered across traces. (Launched in early access in May 2026 and rolled toward general availability over the following weeks.)
- Canvas, rebuilt — a collaborative workspace that combines a chat interface with an autonomous investigating agent, plus Canvas Skills for reusable investigation steps.
- Auto-investigations — automatic data-gathering and a proposed fix the moment an alert fires.
- OTel GenAI-native — the whole thing rides on the v1.40.0 GenAI semantic conventions, which is why there's no lock-in and no re-instrumentation churn.
Pricing#
Honeycomb's free tier includes 20M events per month — genuinely enough to instrument and debug a real production side project before you pay anything, and notably more generous than the free tiers of the big APM incumbents. Paid plans scale on event volume from there.
The honest caveat#
Honeycomb is an observability tool, not an eval harness. It will show you, in exquisite detail, what your agent did and how it collided with the rest of your system — but it won't score whether the answer was good. For output quality you still want an eval layer (Langfuse vs Laminar vs Arize Phoenix covers that decision). The strongest setup pairs the two: evals for "was it right," Honeycomb for "what actually happened in production."
Once the spans are landing, the skill that pays off is reading them. We wrote the companion guide: How to Debug a Multi-Agent Workflow.
