Research A research lab and applied practice

Systems Shape Behavior: Designing AI Around the Outcomes We Want

AI affordances are not fixed. We research multi-agent deliberation, cognitive bias, and architectural design to build systems that actively improve how organizations decide and work.

Read the research → Bring us your question
Five findings · intellectual grounding in 60 seconds

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

01
Deliberation architecture shapes outcomes.
How AI agents are structured, not which model is used, determines recommendation quality.
02
Role framing changes agent behavior.
Using identical evidence, an AI given a CFO persona updates its opinions differently than one given an engineer persona.
03
Aggregation protocols change who "wins."
Somewhere in the system, a rule decides whether a hesitant opinion counts as much as a confident one. Change that rule and a different name comes out on top.
04
AI-assisted decisions embed hidden governance choices.
Most organizations deploying multi-agent AI haven’t mapped where human authority actually sits in their process.
05
Organizational conditions matter as much as the model.
AI performs differently inside different decision architectures. Designing those conditions is the leverage point.

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

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

The question

When several AI agents evaluate the same candidate, what decides the answer? The paper tests whether the method of combining their opinions changes the outcome.

The framework

Three agents, each with a distinct Mental Model of Self, deliberate over ten candidates across five rounds. Sentiment analysis and a non-Bayesian update track how every opinion moves. Three aggregation protocols then combine the same opinions three different ways.

Key findings

The aggregation rule changed the outcome. Six of ten candidates shifted position between Borda count and a conviction-weighted ranking, including the top choice. A third protocol placed seven candidates in one undifferentiated tier and recommended nobody.

Two ways of combining the same three opinions. Two different top candidates.

Ten candidates, one simulation run. Six changed position.

What it means in practice

Any system that pools several AI judgments contains a voting rule. Borda count treats every ranking as equally certain. The conviction-weighted protocol divides each final sentiment by how much that agent wavered, so a confident opinion outweighs a volatile one. Applied to the same conversation, the two rules hand you different names.

When a hiring committee convenes, someone says "let's vote" or "I want everyone comfortable with this." The aggregation is a decision people argue about out loud. Deploy a multi-agent system and that same decision sits in a default nobody experienced as a decision.

Read the full paper · openreview.net →
For your firm

If you use AI for evaluation, scoring, or selection in hiring, vendor review, or strategic prioritization, your system already contains a voting rule. Name that rule, then test it against one alternative on evidence you already have.

Build it into a Tool → · Make it a Product →

Active lines of inquiry

Three open questions shaping current work.

These are live problems we're investigating. We're always adding more. 

01 In progress
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? It changes depending on whether agents are advisory, evaluative, or executive as well as on where humans engage.

For your firm

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

See the Playbook →
02 In progress
Can an AI advisory panel be designed to produce better strategic guidance?

The Artificial Board of Advisors concept asks this directly. An AI panel drawing from behavioral economics, systems thinking, and organizational psychology, may offer new, adversarial counsel where outside perspectives are hard to get.

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

Role Diagnostics maps human-AI collaboration structures. Sometimes humans set the goal and AI executes; sometimes AI surfaces options and humans decide; sometimes both deliberate. Which arrangement wins depends on the problem.

For your firm

This informs how we design every internal Tool. We configure human-AI interaction to the specific problem your team is solving.

See the Tool →
We Are All Agents · newsletter & podcast

Research findings, written for leaders deciding now.

Each issue starts with a research question and ends with an implication you can act on this week. Subscribe →

Applied R&D partnerships

Your question becomes a research question we investigate together.

We take on a small number of applied R&D partnerships each year. Each effort produces a working solution and a rigorous account of what we found. If you have a question to explore in both a scientific and actionable manner, let us know.

Message us on LinkedIn → Or email [email protected]

First conversation is always free. We'll tell you honestly whether your problem is a fit.