LoginSign Up
← Back to all sparks
H

Honeycomb

INFRA · APIS
Velocity7.5

Honeycomb turns agent cost and failures into a dashboard, not a query.

observabilityllm-observabilityagent-monitoringcanvas-agentsmcpcost-tracking
◆Current state
Honeycomb spent October and early November building Canvas, its investigation agent, with connectors, memory, sections and scoped MCP access. On 10 November it made AI Ecosystem generally available: a view of conversation volume, failure rate, token use, latency and estimated LLM spend per agent and model. From any number, users can open the conversation behind it, and an Agent Timeline generates notes on failures, retry loops and cost.
◆Where it's heading
Honeycomb is working on agents from both sides: Canvas uses an agent to investigate systems, and AI Ecosystem observes the agents customers run. Both lean on the same idea of going from an aggregate to the exact trace without writing a query.
◆Prediction
Expect Canvas to start reading AI Ecosystem data, so the investigation agent can explain another agent's failures or cost spikes on request.

◆Recent moves

  1. 3d ago

    AI Ecosystem: now GA

    ⚡ SPARK

    AI Ecosystem goes GA with fleet health, estimated LLM spend by agent and model, and auto-generated timeline notes on failures and retries. It extends Honeycomb from using an agent (Canvas) to monitoring the agents its customers ship.

    View source ↗
  2. 7d ago

    Dark mode settings moved to the Account menu

    The theme switch moves to the Account menu with a System option, and the menu is regrouped. A settings tidy-up between the Canvas releases.

    View source ↗
  3. 9d ago

    Canvas Sections and a better mobile experience

    Sections turn a Canvas investigation into summary cards with the question, evidence, conclusion and next steps, and a clicked next step goes straight to the agent. Mobile support makes reading and steering an investigation possible from a phone, which matters for on-call use.

    View source ↗
  4. 10d ago

    Honeycomb MCP: granular resource access

    MCP connections can now grant read or write access per resource, such as boards, SLOs and triggers, instead of all-or-nothing. It is the permission layer that write-capable agents needed, a day after Canvas connectors gained write actions.

    View source ↗
  5. 10d ago

    Memory in Canvas: now in beta

    Canvas keeps personal memories about your systems across investigations, editable in a Memory panel and readable by MCP clients. It reduces repeated setup in each investigation, though memories stay per-user rather than shared with the team.

    View source ↗
  6. 11d ago

    Reply to selected text in Canvas

    Users can select text in a Canvas agent reply and ask about it directly, with the selection attached as context. It is a small chat UX gain that makes longer agent investigations easier to steer.

    View source ↗