Kitaru
Kitaru is an open-source platform from the ZenML team for making AI-agent runs durable, replayable, and testable. Its newer replay-based evaluation workflow turns production traces into replayable scenarios, cohorts, evaluators, and regression experiments so teams can safely improve agents.
At a Glance
- Engineering teams operating AI agents in production
- Python engineers building long-running or asynchronous agents
- Companies with real user traffic and a need for regression testing
- Teams using coding agents and background/ambient agents
- +2 more
AI Tools by Kitaru
(1)
AIE Talks
AI Engineer Talks Summarized
Discussions
No discussions yet
Be the first to start a discussion about Kitaru
Latest News
Kitaru team published its evaluation of sandbox runtimes for long-lived coding agents.
Introducing the new Kitaru: from production traces to repeatable evals.
Kitaru documented a durable runtime around Claude Agent SDK invocations.
Kitaru documented durable execution for LangGraph agents.
Products & Services
Apache-2.0, self-hosted infrastructure for durable execution of Python agents. It provides flows, checkpoints, persistence, crash recovery, human/agent/webhook waits, replay, a server, workers, CLI, Python SDK, TypeScript SDK, and dashboard.
A workflow that imports or records agent sessions, identifies failure modes, organizes sessions into cohorts, captures expert judgments, creates evaluators, and replays experiments against agent changes.
A hosted version of Kitaru available through cloud.kitaru.ai; the official pricing page describes Kitaru as a flat $39/month plan.
A Kitaru-team-built archive of AI Engineer talks with summaries, quotes, timestamps, searchable talks, and curated packs.
Market Position
Kitaru positions itself as the execution-and-replay layer beside, rather than a replacement for, observability tools and agent frameworks. Compared with trace/observability products such as Langfuse, LangSmith, Braintrust, Arize Phoenix, and W&B Weave, it makes runs executable and replayable; compared with durable-execution/orchestration products such as Temporal, Inngest, Trigger.dev, Restate, Hatchet, and LangGraph, it emphasizes Python-first agent execution plus production-trace replay and regression evaluation. It explicitly says it is not an agent framework or a tracing dashboard.
Leadership
Founders
Adam Probst
Co-founder and CEO of ZenML and a co-founder of Kitaru; Crunchbase identifies his prior co-founder role at MAIoT.
Hamza Tahir
Co-founder and CTO of ZenML and a co-founder of Kitaru; Crunchbase identifies his prior co-founder and CTO role at MAIoT.
Executive Team
Adam Probst
Co-founder and CEO, ZenML (the company behind Kitaru)
Co-founder of ZenML and Kitaru; previously co-founder at MAIoT according to Crunchbase.
Hamza Tahir
Co-founder and CTO, ZenML (the company behind Kitaru)
Co-founder and CTO of ZenML and Kitaru; previously co-founder and CTO at MAIoT according to Crunchbase.
Founding Story
ZenML users began stretching pipeline DAGs to run agents that operated for hours or days, with dynamic branching, durable state, and recovery needs that did not fit a static pipeline abstraction. The team created Kitaru as a separate interface and product for asynchronous, long-running agents, first emphasizing durable execution and then rebuilding it around the observation-to-replay-to-evaluation loop: production traces become safe, repeatable tests for future agent changes.
Business Model
Revenue Model
Open-source self-hosting is free; hosted Kitaru is sold as a flat monthly managed-service plan, with enterprise access through ZenML Pro / a custom plan.
Pricing Tiers
Self-hosted Kitaru, available under Apache 2.0.
Flat monthly pricing according to ZenML's official pricing page.
Enterprise pricing is handled through ZenML / ZenML Pro.
Target Markets
- Engineering teams operating AI agents in production
- Python engineers building long-running or asynchronous agents
- Companies with real user traffic and a need for regression testing
- Teams using coding agents and background/ambient agents
- ML, platform, and MLOps teams managing agent infrastructure
- Organizations that require self-hosting, data control, and infrastructure portability
- Regression testing production customer-support and business-process agents
- Debugging and reproducing failures in long-running or long-horizon agents
- Testing prompt or model changes against real historical sessions
- Reducing token and compute waste by resuming from checkpoints instead of restarting
- Running background coding agents and other ambient agents away from local laptops
- Human-in-the-loop approvals and pauses in agent workflows
- ZenML's Kitaru launch material uses a returns-resolver/support-agent workflow as its worked example, including imported Langfuse sessions and a hazmat-flag failure cohort.