The BraveLabs Lab

The design of an AI system
shapes the decisions it produces.
We study how.

BraveLabs is a research lab and applied practice. We investigate multi-agent AI deliberation, cognitive bias in AI systems, and how AI architecture shapes organizational outcomes. We publish what we find and apply it in practice.

Five findings — for the leader who wants intellectual grounding in 60 seconds

What the research shows about AI, organizations, and how they interact:

These findings inform every engagement we run. See how they apply
What we study

When multiple AI agents evaluate a candidate, assess a strategy, or review a proposal together, the structure of that deliberation matters. Which agents participate. How they update their opinions. How their individual assessments get aggregated into a recommendation. These design choices produce measurably different results from the same underlying evidence.

We study that phenomenon systematically. Our research examines how AI systems form opinions, where cognitive biases enter the process, and how deliberation architecture shapes the quality of collective decisions. We work in organizational settings where design choices carry real consequences: hiring, evaluation, strategic assessment.

Every question we investigate connects to something an organization has to decide.

What this means for your firm
The AI systems you deploy are making choices about how to evaluate, recommend, and decide. Understanding that architecture — and designing it intentionally — is what separates organizations that get lasting value from those that get inconsistent outputs.
Start with a Playbook
Published research

Sentimental Agents: Exploring Deliberation, Cognitive Biases, and Decision-making in LLM-based Multiagent Systems

Elizabeth A. Ondula  ·  Daniele Orner  ·  Nick Mumero Mwangi  ·  Casandra Rusti

Fourth Workshop on Knowledge-infused Learning, IJCAI 2024  ·  Barcelona, Spain

The question

How does sentiment shape opinion dynamics in multi-agent AI systems — and what cognitive biases emerge when those systems evaluate people?

The framework

Sentimental Agents gives each AI agent a distinct Mental Model of Self. Sentiment analysis and a non-Bayesian update mechanism track how each agent's opinion forms and shifts across rounds of deliberation. The framework was applied to a simulated HR recruiting environment, where three agents in distinct organizational roles evaluated ten candidates and produced collective recommendations via three decision protocols: Borda Count, Tiered List, and Gut Feeling.

Key findings

Agents showed a modest positivity bias — the positive sentiment slope ran approximately 29% steeper than the negative. Drift varied by agent role: the VP of Engineering and Plant Manager adapted more dynamically across the conversation than the CFO, suggesting role framing influences how a Mental Model of Self responds to deliberative pressure. Different aggregation protocols produced different candidate rankings from identical evidence.

What it means in practice

Deliberation architecture shapes the recommendation. The number of agents, how they update, how their outputs are combined: these are design variables with measurable consequences. Organizations designing AI-assisted evaluation systems should treat these choices with the same care as model selection.

What this means for your firm: If you're using AI for evaluation, scoring, or selection — hiring, vendor review, strategic prioritization — the design of your system is producing different outcomes than you think. The architecture is the decision.

Active lines of inquiry

Three open research questions shaping current work. Not papers-in-progress — live problems we're investigating.

01

How does multi-agent design distribute authority in organizational processes?

When a multi-agent system is embedded in a workflow, a governance question follows: who is making the decision? The answer changes depending on whether agents are advisory, evaluative, or executive, and where humans engage in the process. Most organizations deploying multi-agent AI today have not mapped their system against this typology. We are developing the framework.

In progress
What this means for your firm

Before you can optimize AI-assisted decisions, you need to know who's actually deciding. A Playbook makes that mapping explicit.

See the Playbook →
02

Can an AI advisory panel be designed to produce better strategic guidance?

The Artificial Board of Advisors concept asks this directly. An AI advisory panel drawing from behavioral economics, systems thinking, and organizational psychology offers a form of adversarial counsel that human boards rarely deliver consistently. We are investigating whether intellectual diversity in the panel architecture changes recommendation quality, and how that diversity can be designed.

In progress
What this means for your firm

This is the research behind our Playbook design. Better AI advisory structure produces better strategic decisions from the same leadership team.

See the Playbook →
03

Which human-AI configurations produce better decisions for which problems?

Role Diagnostics maps human-AI collaboration structures against the kinds of decisions organizations face. In some configurations, humans set the goal and AI executes. In others, AI surfaces options and humans decide. In others, both deliberate across multiple rounds. The research question is which arrangement produces better outcomes for different problem types.

In progress
What this means for your firm

This research directly informs how we design every Tool — matching the human-AI configuration to the actual problem your team is solving.

See the Tool →
We Are All Agents

Our newsletter and podcast

We Are All Agents surfaces findings from our own research and from across the organizational AI research community. Each issue is written for leaders making real decisions about AI now. It starts with a research question and ends with an implication you can act on.

Every issue connects research findings to something your organization can decide or act on this week.

See how research shapes our engagements →
Recent issues
Why 70% of Leaders Are Ready to Hand Their Decisions to Robots
From Email to AI: The Hidden Pattern Behind Tech Adoption Failure
Ecosystems, Travel Nurses, and the Future of Employment
Applied R&D partnerships

Your question becomes a research question we investigate together.

BraveLabs takes on a small number of applied R&D partnerships each year. Your organizational AI challenge becomes a research question we investigate together. The output is a working solution and a rigorous account of what we found.

You don't just get a consultant. You get access to the only firm that's actively studying what you're trying to solve.

Areas of particular interest

Evaluation architectures — how AI systems score, rank, and select

Multi-agent deliberation — structured collective AI reasoning

Decision support systems — AI-assisted strategic and operational judgment

Researchers
Principal Investigator
Elizabeth A. Ondula
University of Southern California
Study design, analysis, and primary authorship. Research focus: multi-agent systems and reinforcement learning.
Daniele Orner
Co-founder, BraveLabs
Framework design, systems architecture, analysis, and writing. Research focus: cognitive models, NLP, and multi-agent deliberation.
Ibanga Umanah
Co-founder, BraveLabs
Research questions, study design, and writing. Focus: organizational science and applied AI.
Nick Mumero Mwangi
BraveLabs
Research infrastructure, implementation, and data engineering.
Academic Advisors
Joshua Becker
University College London
Multi-agent models and group decision-making.
Peter Law
Deliberative facilitation and group dynamics.
Work with the lab

The research is the practice.
Bring us a real question.

We take on a small number of applied R&D partnerships each year. In the first conversation, we'll tell you honestly whether your problem is a fit — and what working together would look like.