Honeyhive

HoneyHive operates an agent observability and evaluation platform for teams running AI systems in production. The product covers three workflows: observation (distributed tracing that replays each agent run step by step across tool calls, prompts and decisions, including long-running trajectories spanning hours or days with retries, loops and sub-agent handoffs; live-traffic monitoring; and alerting with thresholds, drift detection, email and webhook notifications), evaluation (experiments comparing agent versions on the same dataset, regression tracking, datasets curated from production failures, LLM-as-a-judge and code-based evaluators scoring outputs, tool calls, agent-to-agent interactions and full trajectories, plus human review through annotation queues and custom rubrics), and improvement (root-causing failures, versioning prompts, re-running evaluations, and gating releases with automated evaluations in CI).

The platform is built on OpenTelemetry standards so that it integrates with a team's existing stack. It offers Python and TypeScript SDKs with automatic instrumentation for more than 50 libraries, including LangChain, LangGraph, AWS Strands, Google ADK and the OpenAI Agents SDK; other languages can send OpenTelemetry traces to a HoneyHive OTEL collector or use its APIs. Additional components include prompt management (playground, versioning and history, prompt deployments, custom model providers, functions and external tools), a CLI, an OpenAPI spec, a daemon for Claude Code telemetry and coding-agent exporters, an n8n ETL integration, and MCP-accessible "Skills" that let coding agents fetch alerts, traces and evaluation context. HoneyHive positions itself as a collaborative system of record that versions traces, prompts, tools, datasets and evaluators for auditability and compliance in regulated environments.

Founding story

The company was founded by CEO Mohak Sharma and CTO Dhruv Singh, motivated by a pattern they observed across hundreds of teams: AI prototypes were quick to build but broke in unexpected ways in production because non-deterministic LLMs defeat traditional unit and integration testing, black-box models offer little visibility for debugging compound systems, and AI development does not fit the traditional software development lifecycle.

Business model

B2B software sold as a hosted platform with a free self-serve Developer tier and a custom-priced Enterprise tier. Usage is metered in "events," defined as trace spans plus metric-label combinations sent to the API via OTLP or JSON.

The Developer plan is free (no credit card required) with 10,000 events per month, up to 5 users, one workspace, 30-day retention and community support. The Enterprise plan is custom-priced and adds custom usage limits, unlimited users and workspaces, SAML and custom SSO, custom roles and permissions, custom retention and data residency, uptime and support SLAs with service credits, custom DPA and BAA, a dedicated technical account manager with QBRs, and optional self-hosted, hybrid or single-tenant deployment. Discounts are offered to startups that have raised under $5M in total funding.

Traction

During its 2024 beta the platform grew more than 50x in requests logged and the company doubled its team size, with early customers spanning AI startups and Fortune 100 enterprises in insurance and financial services. Reported customer outcomes include a 340% improvement in one customer's agent accuracy on web-browsing tasks and a 5x faster development cycle at another enterprise. HoneyHive states it powers observability and evaluation across dozens of mission-critical AI applications at Commonwealth Bank of Australia, covering agents that serve 17M+ retail consumers and 55K internal users.

Latest developments

The current product line emphasizes long-running agent debugging, live production scoring with intelligent sampling, human review queues, CI release gating, and integration with coding agents such as Claude Code via an MCP server, a telemetry daemon and purpose-built Skills. Public repositories updated in 2026 include a CLI, control-plane and data-plane TypeScript SDKs, a Homebrew tap, a Claude Code telemetry daemon, an n8n shipper and a skills repository.