ML Research Intern in Agentic Runtime Systems

The frontier models are converging into commodities; the engine is no longer the differentiator. The scarcity in 2026 is not the model. We are building the runtime systems that turn probabilistic token generation into reliable workers that finish 50-step tasks without spiraling. We are solving the math of multi-step reliability through engineering, not wishful thinking.

Responsibilities

Work on the Agentic Harness, treating the LLM as a component within a larger, deterministic control system; Architecting loops that survive error accumulation; Build verification layers that force the agent to recognize its own hallucinations and self-correct; Implement systems where the agent compiles its own constraints and the harness enforces them via constrained decoding or external execution; Optimise KV-cache hit rates and design restorable compression strategies to prevent attention decay

Requirements

We are looking for systems thinkers who happen to know ML, not the other way around; You know that a demo is not a product and are interested in the boring engineering: state machines, retries, idempotency, serialisation; You are comfortable working with the raw output of the model and building robust software in Python or Rust around it; You understand why dynamic tool loading destroys inference economics and prefer deterministic guardrails over prompting; You are willing to read a 500-line trace of a failed agent loop to find the exact moment the context drifted; You care about the last mile of reliability