Hamza Tahir is a ZenML cofounder building infrastructure for machine-learning and agent systems. His Kitaru work connects production traces with evaluation, replay and the execution environment needed to test a change.[2][1]
Start with Kitaru →
Education
Program attendance; Fall 2017 class (award unspecified)
[6]Links
Current professional profile ZenML founder record Cofounder · ZenML
Career
Tahir’s owned developer biography describes a path across software engineering, machine learning, data engineering and product work. It names study at the Technical University of Munich. CDTM lists him among MAIoT’s founders in its Fall 2017 class. ZenML’s own team record identifies him as cofounder alongside Adam Probst; Kitaru extends the company’s infrastructure work into production agents.[3][6][2]
Contributions
Turn observed sessions into a tested intervention
Current Kitaru documentation starts with importing recorded agent sessions, finding patterns across a cohort and turning a finding into a check. A replay experiment tests a proposed fix through a worker installed beside the project’s dependencies. The local setup uses Docker or Podman, while model and tool access remain the operator’s responsibility. This connects analysis of past behavior with an executable comparison instead of stopping at a trace display.[5]
Compare decisions after changing one condition
In his save-button talk, Tahir demonstrates checkpointed execution that can be replayed with a changed model, prompt or tool response. Comparing the fork with the original run exposes differences in decisions, cost and divergence. He then applies an intervention across a selected cohort of production cases. A cheaper fork is useful evidence only when the resulting behavior also meets the service’s requirements.[4]
Tools, writing & talks
Tools
Start here
Kitaru
github.comOpen-source infrastructure for inspecting agent sessions, creating checks and testing changes through replay.
Read session import and local-worker setup, then test a small recorded task with a known expected result. Python and terminal familiarity help; local serving needs Docker or Podman, and replays need the project’s model/tool access.
Writing, research & talks
Hamza Tahir argues the most underrated primitive in agent infrastructure is a save button: freeze an agent to durable state while it waits on a tool or a human, restore in milliseconds, and afterwards reload any point in its trajectory to test a different prompt, model or tool.
Catalog workOriginal source- Your Agents Need a Save Buttonai.engineer
A demonstration of checkpoint replay, targeted interventions and cohort-level evaluation.
Presenter[4]