▶ Watch ↗17:13
Conference affiliation: Head of Deployed Engineering · Factory · 2026
Nikhil Harithas works on the infrastructure and feedback that let organizations put AI agents to work. As Factory’s head of deployed engineering for its 2026 AI Engineer appearance, he explained how coding agents could run on customer-controlled machines with deliberately scoped permissions. His work at Factory and earlier at Patreon makes those deployment questions concrete: where computation runs, which credentials software receives, and how engineers check its output.
Before Factory, Nikhil was a senior machine-learning engineer at Patreon, where the team needed to transcribe creators’ audio and video and provide automatic captions for its native video product. Serving Whisper involved tradeoffs among cost, infrastructure work, and data security. The team adopted Baseten to serve Whisper, reducing the infrastructure burden so it could concentrate on the product. Nikhil emphasized deployment speed, cost, and compliance alongside the model’s capabilities.
At Factory, Nikhil helped build a data function around dbt and Droid, the company’s software-development agent. His account of that work describes a change in approach as the system grew. Rebuilding entire tables had initially been adequate. As queries became more expensive and the team wanted fresher data, it moved toward incremental builds, updating only the relevant portion of a table. That made assumptions such as each row having a distinct identifier consequential: the system needed to identify which records to update. The team began enforcing those assumptions with tests, starting with its most important tables.
The instructions supplied to Droid created another problem. Adding more skills—reusable instructions for particular tasks—could make it harder for the agent to determine which one applied. Nikhil favored fewer, denser skills, explicit pointers in AGENTS.md, and repository organization that made relevant instructions easy to find. Tests and discoverable instructions gave the agent concrete ways to check its work rather than relying on a larger collection of guidance.
Nikhil’s explanation of Factory’s deployment design addresses customers whose requirements extend beyond choosing a model. Blackstone wanted control over compute, inference, and identity. Factory’s response was a team effort; Nikhil explained how its architecture separated decisions that a hosted coding agent might otherwise bundle together:
These distinctions explain what customer-controlled deployment requires in practice. Choosing a machine does not settle which identity an agent uses, and preserving a task does not require keeping its original machine alive.
▶ Watch ↗17:13
Loading more talks and highlights…