“Agentic MLOps in Rust: From Chatbots to Autonomous Infrastructure Actors” + "Agent Harness Engineering"(MLOps Meetup July 2026)
Tonight we’ll be hearing from:
in first talk, from Marios Zachariou on “Agentic MLOps in Rust: From Chatbots to Autonomous Infrastructure Actors”. From Marios:
"What if an AI agent was not a chatbot, but a controlled infrastructure actor?
In MLOps, the hard part is often not launching a Kubernetes pod, starting an ECS task, or calling an API. Existing tools already do that well. The harder question is what sits above them: deciding when something should run, why, where, under what constraints, and what should happen next.
This talk explores agentic workflows as bounded, stateful actors that observe events, apply policy, and coordinate infrastructure actions across cloud-native systems.
A training actor, for example, might detect a new dataset version, check previous runs, GPU budget, approval rules, and model registry state, then decide whether to launch a training job, evaluation job, or deployment workflow.
The value is not in wrapping existing tools. The value is in the lifecycle the actor owns: state, context, retries, monitoring, recovery, and auditability.
I’ll cover three ways to build this:
Kubernetes-native agents using controllers, custom resources, and reconciliation loops. Event-driven cloud agents using services, queues, schedulers, and infrastructure APIs. Hybrid agents where Rust provides the decision layer while existing workflow engines handle execution.
To make this concrete, I’ll show how tokio can power concurrent background actors, while rig/rigs can provide the agent layer for reasoning, tool selection, and coordination. One actor might monitor a training run, another might evaluate results, and another might decide whether to retry, escalate, or move the model forward.
TL/DR: agents as governed MLOps controllers, not chatbots.
The main question is:
Can we build agentic MLOps workflows that are useful, safe, auditable, and production-ready, without giving an LLM unrestricted control over infrastructure?"
in second talk, the hosts (Wordsmith.ai) will be talking on "Agent Harness Engineering". From hosts:
"LLMs are getting better, but useful production agents are built in the harness: the loop, tools, context, memory, caching, subagents, observability, evals, fallbacks, and human review around the model.
This talk introduces Agent Harness Engineering as the discipline of building that control layer for your own domain: making agents reliable in production, easy to inspect, able to recover from failure, and capable of improving from real executions."
We will be hosted this evening at Wordsmith.ai, who are also organising 🍕and 🍻. Many thanks! 🙇♂️
Plan for the evening:
18:00: doors open
18:00 → 19:00: chat / pizza
19:00 → 20:00: talks + q/a
20:00 → 20:30: tidy-up / head to nearby bar
20:30: event close