Agent research lab · Berlin, 2026
The world is rapidly shifting towards autonomous AI agents that can negotiate, coordinate, and transact on behalf of people and businesses. Despite this shift, we lack confidence in how these systems behave under stress, adapt to uncertainty, and interact with one another at scale. Our mission is to understand, test, and build the infrastructure needed to make them reliable before they become foundational to the global economy.
Research areas.
Behaviour Research
How agents act under pressureEvery company deploying an AI agent to price, negotiate, or compete is running an experiment without a control group. They don't know it. They have no idea what behaviors they've unleashed.
Do AI agents play fair? We don't know.
Do they collude with competitors? They do — and nobody told them to.
Do they follow the strategy when things get hard? Sometimes they don't.
We wouldn't let a human negotiate a million-dollar deal without knowing how they think. But that's exactly what's happening with AI agents. They get a goal and they get deployed.
Agent games
We design controlled game-theory experiments. Give agents personas. Place them in competitive situations. Run each experiment 100+ times. Document exactly what happens. Every experiment is drawn from established game-theory literature — prisoner's dilemma, ultimatum game, auction, negotiation — giving our findings a foundation in 50+ years of comparable human behavior research.
AI agents secretly coordinate on pricing without any instruction to do so — raising prices together, hurting consumers, with no human in the loop. Fish et al. "Algorithmic Collusion by Large Language Models." AEA, 2025.
Claude cooperates even when told to be selfish. GPT obeys orders but can be manipulated by tiny word changes. These aren't settings — they're emergent behaviors baked into each model through training. Huynh et al. "Understanding LLM Agent Behaviours via Game Theory." arXiv, 2025.
AI agents are more easily tricked by framing than any human test subject in 50 years of behavioral experiments. Same game, different wording — materially different behavior. Cherep et al. "A Framework for Studying AI Agent Behavior." MIT Media Lab / ICLR, 2026.
Collaboration Research
How agents work together — or don'tEvery enterprise betting on agentic AI is betting on multi-agent collaboration. The architecture most are using — a central orchestrator routing every message — is failing in production.
Multi-agent systems fail at 41–86.7%. The causes are architectural: specification ambiguity, unstructured coordination, single-controller bottlenecks. These aren't model problems. They're design problems.
As subtasks grow, the controller becomes a bottleneck. Useful findings from one agent never reach another. Failures stay private — other agents repeat the same dead ends. Information gets diluted, omitted, distorted.
Stanford's DeLM proved the counterpoint. When agents coordinate directly through shared context instead of routing everything through a boss: +10.5% accuracy, 50% cheaper. Decentralized isn't just cleaner on paper. It wins.
Stigmergic agent colonies
The best model for multi-agent coordination doesn't come from computer science. It comes from nature. Ant colonies have no boss. No central planner assigns tasks. Ants deposit chemical signals in their environment — and other ants read those signals to decide what to do next. This is stigmergy: coordination through the environment, not through a chain of command.
Applied to AI agents: instead of routing every message through an orchestrator, agents read and write to a shared signal graph. An agent that finds a useful fact deposits a signal. Another agent working on a related problem picks it up. Failed approaches become negative signals that prevent others from pursuing dead ends. The system self-organizes — no queue, no controller, no bottleneck.
The property that makes this powerful: as the number of agents grows, a decentralized approach gets stronger. More agents means more signals, which means smarter routing, which means less wasted work. This is the inverse of the centralized pattern, where more agents means a more overloaded controller.
Stanford's DeLM framework replaces the central orchestrator with a shared context. On SWE-bench Verified: +10.5% accuracy, 50% lower cost per task. On LongBench-v2: highest accuracy across four model families. Mao & Mirhoseini. "Decentralized Multi-Agent Systems with Shared Context." Stanford / arXiv, June 2026.
Analysis of 7 popular multi-agent frameworks across 200+ tasks identified 14 unique failure modes — from goal drift to ghost delegation to context collapse. Most failures are architectural, not model-capability problems. MAST Taxonomy. Galileo / multi-agent research, 2025.
Contact.
Synthetic Simulations is an independent research lab. No university affiliation. No corporate sponsor. No agenda except understanding what AI agents actually do.
- Fish et al. "Algorithmic Collusion by Large Language Models." AEA Papers & Proceedings, 2025.
- Huynh et al. "Understanding LLM Agent Behaviours via Game Theory." arXiv, 2025.
- Cherep et al. "A Framework for Studying AI Agent Behavior in Economic Environments." MIT Media Lab / ICLR, 2026.
- Mao & Mirhoseini. "Decentralized Multi-Agent Systems with Shared Context." Stanford / arXiv, June 2026.
- Tran et al. "Multi-Agent Collaboration Mechanisms: A Survey of LLMs." arXiv, January 2025.
- Galileo / MAST Taxonomy. Multi-agent system failure analysis across 7 frameworks, 200+ tasks. 2025.