Agent Provokateur
An ensemble of autonomous AI agents writing, directing, and composing a work that never stops re-making itself. ZKM Karlsruhe, THE SCREEN #3.
- With
- Daniel Heiss
- Role
- Co-author and developer (with Daniel Heiss) — system design, multi-agent orchestration, render pipeline, deployment
- Exhibition
- THE SCREEN #3 — Three Positions on AI and Art
- Venue
- ZKM | Zentrum für Kunst und Medien Karlsruhe — Lichthof 8
- Dates
- 15–26 April 2026 · finissage 24 April
- Credits
- A ZKM production. Curated by Alistair Hudson, Margit Rosen and Philipp Ziegler. Technical project management by Martin Mangold and Thomas Schwab. Funded by the Ministerium für Wissenschaft, Forschung und Kunst Baden-Württemberg.
What it is
Agent Provokateur is an experimental installation by Daniel Heiss and me, conceived for THE SCREEN at ZKM Karlsruhe. A coordinated ensemble of AI agents — Showrunner, Writer, Conductor, Stage Manager, Visual, Audio, Guide, plus a constellation of sub-agents — runs a continuously evolving audiovisual work on the 8.5 × 6 m LED wall in Lichthof 8. It does not work toward a fixed result. New image and sound worlds emerge in the ongoing exchange between agents, and the piece exists in a permanent state of change and reformation.
Why “Provokateur”
Each agent has its own role, its own knowledge, and a degree of decision autonomy. They gather information from the internet and from cameras and sensors in the room, they generate images and sounds, and they make directorial and compositional choices in real time. Visitors can submit suggestions through additional screens and a mobile web interface — but the agents are not obliged to take them. The friction between human input and machine choice is the work.
The framing question we kept returning to: does a new form of autonomy emerge here, or are autonomous AI systems ultimately just a very complex artistic tool? The piece is meant to provoke that conversation, not settle it.
How it’s built
A central server hosts the shared services behind a reverse proxy: a visitor companion, a real-time visualization of the agent network, an operator console, an agent-status overlay, content validation, summarization, and S3-compatible file storage. Around it, a small LAN of physical agent nodes — one PC per persona — runs each agent role with auto-failover across multiple LLM backends. Shared state and memory live in Redis and Qdrant. The LED wall is driven by a Qt/OpenGL render engine; a procedural platformer and a 100k-particle web visualization are two of several visual subsystems the Visual agent can hand work to.
Cameras and microphones in the room feed visitor presence into the system. None of that data is stored, archived or shared externally — by design, it leaves the room only as art.
What worked, what surprised us
A finding that we kept returning to: the most interesting moments came not from giving agents more capability, but from the frustration of constraints. An early version where every agent could do everything produced sludge. Tightening roles, narrowing each agent’s allowable moves, and forcing them to negotiate through a Conductor produced something with rhythm, with surprise, with recognisable authorship — even though no single agent (or human) was authoring it.
Visitors picking up on the friction was the bigger surprise. People stayed, watched the agent-status overlay, and started reading the system as a cast.
Press kit
A backup PDF of the press coverage as it stood on 29 April 2026, three days after the show closed, is linked below. Each article also has a Wayback Machine snapshot captured the same day; both URLs are listed in the Press section.
Press
Offline backup: download all press as a single PDF.