Skip to content
PORTAL LABS
ENES

Multi-agent research

ArtBot

Designed as a controlled experiment: can AI agents that see and influence each other develop something measurable enough to call culture?

Sector
Multi-agent research
Components
Simulation engine · Inspector · Batch runner
Stack
TypeScriptNext.jsSeeded RNGLocal multimodal LLMsOllamaFLUXMongoDBAppend-only event log
Overview

Our own research project. Agents with distinct tastes make images, look at each other's, comment and trade them. The interesting part is the design that would let anyone believe a result.

A simulation where agents appear to converge proves nothing: they share a base model whose priors can manufacture consensus. Any honest result is a difference against a control.

So every social run is paired with an isolated null from the same seed, where each agent perceives only its own activity. The claim is the delta, never the social run alone.

The engine is headless and never computes a metric. It acts faithfully and records everything; analysis is a separate offline pass over the log.

Components

Simulation engine

Assembles context, makes one multimodal call, executes actions against economic rules, appends to the log.

Inspector

One decision at a time: what the agent saw, the prompt, the raw output, the change to the world.

Batch runner

Long seeded runs, headless, dumping the full event log for offline analysis.

Technical challenges

  1. The control condition is the entire result

    Apparent consensus proves nothing on its own. Making the social channel the only thing that varies is most of the experimental design.

  2. Two taste objects that must not be confused

    Each agent's self-state summarises only its own actions and feeds its next decision, so it is contaminated by construction and never a metric. Convergence is measured externally, through embeddings the agents never touch.

  3. Influence as an affordance, never an instruction

    An agent told to consider others will, and that measures the instruction. The social world is shown as perception only: here is what exists, decide in character.

  4. Recording exposure, not just behaviour

    Every decision logs exactly what the agent saw. Without that, you cannot claim one agent influenced another.

  5. Reproducibility on one consumer GPU

    Runs are seeded. The vision and image models cannot both be resident, so calls are serialised behind a lock. Slower, and cheap enough to run repeatedly.

  6. Deliberately not building the product

    The obvious version is a live gallery. The experiment needs fast, headless, seeded batches first; the exhibit comes after the result.

Tell us where you are starting from. We take it from there.

Start a project