Stop Building AI Agents: Build Harnesses Instead | Hamza Tahir (ZenML / Kitaru)

Everyone is building AI agents.OpenAI SDKs, Claude Code, Deep Agent systems, custom workflows, and orchestration frameworks all promise more autonomous AI. But as these systems become more capable, they start running into familiar engineering problems:retriesstate managementorchestrationcontext controldurable executionThis week we're joined by Hamza Tahir, CTO and co-founder of ZenML and creator of Kitaru, to discuss what happens when agents stop being simple chat interfaces and start behaving like long-running distributed systems.We explore:what an agent harness actually isdurable execution and why it mattersorchestration vs business logicstate management for long-running agentsretries, checkpoints, and human-in-the-loop workflowscontext management and token costsopen vs closed agent frameworkswhy everyone seems to be rebuilding the same layer of infrastructureOne of the biggest questions we kept coming back to:What is a meta harness?If you have an answer, let us know in the comments.Kitaruhttps://github.com/zenml-io/kitaruZenMLhttps://www.zenml.ioHamza Tahirhttps://www.linkedin.com/in/hamzatahir/Pedro Agentwarehttps://github.com/Soypete/pedro-agentwareOpenAI Agents SDKhttps://platform.openai.com/docs/guides/agentsTemporalhttps://temporal.ioDBOShttps://www.dbos.devApache Airflowhttps://airflow.apache.orgPrefecthttps://www.prefect.ioDomesticating AI is a bi-weekly podcast about practical AI for developers.We help you brace the feral open-source AI landscape — so you can tame it instead of getting dragged by it.Subscribe on YouTube, follow on Spotify or Apple Podcasts, and support the show on Patreon.Keep your AI on a leash.Links