Hamza Tahir says he co-founded ZenML about five years before this episode, when MLOps interest was at a peak.
“So I co founded ZenML about five years ago, so this was really, almost, um, at a point where ML Ops was really reaching fever pitch”
Person
Hamza Tahir
1 episodes · 23 quoted claims
Hamza Tahir says he spent four or five years deploying ML models at a previous company he co-founded before starting ZenML.
“And I had done four or five years of that in my previous job, where I was co founding um, another company, uh, trying to deploy ML models”
ZenML grew out of an internal framework for writing workflows and DAGs, was open sourced, gained early traction and revenue, and raised funding.
“And that turned out to be ZenML. We open sourced it, we got a bit of traction at the beginning and a few projects and and revenue, and we raised”
ZenML introduced a dynamic mode so workflow graphs could be defined at runtime rather than compiled statically in advance.
“So we. We introduced a new dynamic mode.”
Tahir describes a large language model as a token generator that cannot take actions on its own.
“The LLM model is simply a token generator, so in itself it doesn't do anything, it doesn't do action.”
Tahir says action-taking by models began about three years earlier with structured outputs and tool calling.
“So the way we did action was we started three years ago with structured outputs and tool calling”
Tahir defines an agent as the combination of a model and a harness, the program that maps model tokens to actions.
“And the the outcome that comes out of combining the harness and the model is the agent, right?”
Tahir says Claude Code is a harness running on Opus 4.8, and that the model has learned to recognize the harness it runs in.
“CloudCode is the harness. And underlying it is Opus 4.8, and um, Opus 4.8. By now, unlike Opus 3.5, understands what CloudCode itself is”
Tahir says placing GPT 5.5 inside Claude Code would perform worse because models and harnesses have become coupled through reinforcement learning.
“If you drop GPT 5.5 into Cloud Code, the harness, that would not be as accurate, it wouldn't perform as good”
Tahir names Pi, LangGraph and Pydantic AI as open harnesses that argue for standards not tied to a single model.
“So we had Pi, we have frameworks like LangGraph or Pyantic AI, and those harnesses, uh, are of the opposite opinion.”
Tahir says moving agents from laptops to cloud infrastructure is underway in 2026, with adoption described as slow and then sudden.
“it's like this sort of slow adoption, and then suddenly boom, there's like a crazy increase.”
Tahir cites Uber as an example of a company whose strong ML platform helped it win its market, including features like surge pricing.
“Similar to ML Ops, where the people who build the best ML Ops platforms like Uber, they won them, their their markets.”
Tahir describes a standard architecture where an API server places events on a message broker and workers process them, so tasks survive worker failures.
“So a very normal one o one architecture is just putting a message queue, a message broker in the middle of that.”
Tahir says agents that execute code need sandboxes, which explains the popularity of providers such as E2B and Daytona.
“it obviously has to execute code, right? So it needs a sandbox, and this is why the sandbox providers like E2B, Daytona model are so popular nowadays”
Tahir says Kitaru aims to be an open runtime that can deploy any agent harness.
“So our our goal is to build an open runtime that allows you to take any harness, any and and deploy it in a way”
Kitaru's first priority was building adapters for popular harnesses such as the Anthropic Agents SDK and OpenAI Agents SDK.
“how do we get the best world class adapters for all the popular harnesses like Anthropic Agents SDK, OpenAI Agents SDK?”
Kitaru makes agent runs resilient to failed tool calls by storing run state in an external database or blob storage.
“by storing that state in an external database or or blob storage.”
Kitaru shipped with an MCP command line interface on day one so other agents can eventually analyze and optimize runs.
“We shipped an MCP CLI day one of Kitaru because we know that eventually that problem needs to be solved by other agents.”
Tahir advises Kitaru users to run it for a week, then filter for the most expensive successful traces and inspect checkpoints for bottlenecks.
“run it for a week and then after a week, just filter for the most expensive traces.”
Tahir compares the current agent tooling boom to the 2021 MLOps period, when many tools appeared and investor money flowed in.
“when we had an explosion of tools, when we had these ML Ops problems and investor money came in.”
Tahir says the open model GLM reached roughly 95 percent of the performance of Opus 4.8, and that this was about two weeks old news at recording.
“The GLM is as basically 95 of Opus 4.8, which is just phenomenal.”
Tahir cites MiniMax, Gemini and GLM, plus Mistral in Europe, as models making open or alternative options economically viable for enterprises.
“Take min, like Minimax, or or like, take the Gemini models, or or, or, or, or take something like GLM coming out of, um, like China right now”
Tahir credits Peter Steinberger, Mario and Armin with better harness ideas behind Pi and OpenClaw, which became popular around December 2025.
“Peter Steinberger and, uh, Mario and a Armin, like those guys, they had better ideas how to manage that program, which just caught fire.”