Back to Curated
VideoPalantirJul 18, 2026

Palantir Agent Engine: Agents as Distributed State Machines

Palantir launches Agent Engine + Agent SDK at DevCon 6: context items, events, and effects turn agent loops into multiplayer distributed state machines.

This is Palantir's product launch talk from DevCon 6 (about 20 minutes), presented by Group Lead Natasha Armbrust, introducing Agent Engine and the companion Agent SDK — a set of core primitives for developers building agents.

The system is built from three primitives: context items define agent sessions through strongly typed data; events let agents act expressively on changes in state; effects connect agents to the outside world. In Armbrust's words, by stitching these three together you get an agent loop that is actually a distributed state machine under the hood — and multiplayer by design.

The talk lands on a live demo of a patient discharge agent: it pauses for nurse approval, reassesses itself the moment a new lab result comes in, and renders its reasoning clearly to every human in the loop — all in a few lines of code.

If you're building enterprise agent systems that need human-in-the-loop approval, long-running state, and multi-user collaboration, this launch is worth watching in full: it shows what it looks like to design agents as state machines rather than chat loops.

FAQ

What is Palantir Agent Engine?

An agent-building product Palantir launched at DevCon 6, paired with the Agent SDK. It gives developers core primitives so the agents they build form a distributed state machine under the hood and are multiplayer by design.

What are Agent Engine's three core primitives?

Context items (defining agent sessions through strongly typed data), events (acting expressively on changes in state), and effects (connecting agents to the outside world). Stitched together, they form an agent loop that is a distributed state machine.

What does the video demo show?

A live patient discharge agent: it pauses for nurse approval, reassesses itself the moment a new lab result arrives, and renders its reasoning clearly to every human in the loop — all in a few lines of code.

Who is this video for, and how long is it?

About 20 minutes, presented by Palantir Group Lead Natasha Armbrust and published in July 2026. Best for engineers building enterprise agent systems that need human-in-the-loop approval and multi-user collaboration.

View original