Research at Foaster.ai

At Foaster.ai, we develop AI agents every week, constantly pushing their limits.

Our belief is simple: AI agents are becoming digital teammates. As they gain responsibility and autonomy in critical tasks, understanding their behavioral patterns, decisions, and social dynamics becomes essential.

Model Intelligence: aiming for the best agent-model fit

We conduct applied research to map the actual behavior of models, to match the right model with the right agent (sales, support, back-office, monitoring...).

Rather than judging LLMs solely on code or math, we test their social, strategic, and long-term behaviors in multi-agent contexts — which makes agents reliable in the real world.

Why it's decisive

Choosing a model is no longer a matter of brand or specs. It's a question of fit: persuasion vs. resistance, cooperation style, failure modes, latency/cost trade-offs, robustness under pressure.

Our research provides evidence, not guesswork.

Our method, in brief

Hierarchical multi-agent simulations with tools, roles, and incomplete information — much closer to real workflows than static prompts.

• Role-conditioned metrics to analyze models from different angles.

• Behavioral signals beyond win rates.

• Reproducible protocols and agent-with-tools framework for realism.

Focus: the Werewolf benchmark

Why Werewolf

A 100% language game, adversarial and socially demanding: hidden roles, uncertainty, evolving narratives.

It reveals if a model knows how to plan over several days, coordinate, persuade, bluff or withstand pressure — exactly the skills that make enterprise agents robust.

What we're running

Round-robin matches between models, balanced by role, with Elo leaderboards and breakdown by role (wolves = manipulation, villagers = resistance).

We also capture public vs. private reasoning to study intention vs. narrative — how a model wins (or fails) in reality.

What you receive

A model leaderboard and model cards detailing strengths, weaknesses, and failure modes.

Concrete agent-model recommendations (e.g., which model to place behind your outreach agent vs. your monitoring agent).

Guardrails & prompts tailored to the trends of each model, plus budget/latency advice for production.

And then

We move on to longer and more complex games, to more model families, and to expanded behavioral metrics.

The goal is simple, deliberately competitive: who can beat the current leader?

Want to have your model evaluated or co-fund larger runs? Contact us.

Don't choose your model blindly.

We benchmark its behavior as an agent and then integrate the best fit into your stack.