Synod

Inspiration

My sister works in recruitment, where every conversation is a negotiation. She often talked about having to balance persuasion, trust, verification, timing, and long-term relationships all at once. The challenge wasn't generating arguments, it was deciding which signals to trust when they conflicted.

Watching those conversations made us realize that good decisions come from weighing conflicting signals—not eliminating them. Yet most AI agent societies are designed to collapse disagreement into consensus as quickly as possible.

In real decision-making, experts disagree because they notice different risks, incentives, and uncertainties. Collapsing those perspectives too early can throw away exactly the evidence needed to make a better decision.

That made us wonder: what if an AI system didn't force its agents to agree? What if disagreement itself carried information about the decision?

That question became Synod. Instead of collapsing different opinions into one answer, Synod preserves disagreement between independent specialists and lets a deterministic chair resolve the final decision.

What it does

Synod is an AI decision society: five independent specialist agents coordinated by a deterministic chair.

Each agent independently analyzes the same situation from a different strategic perspective. They debate by challenging each other's reasoning, but they never vote a winner into existence.

Instead, the chair reads the complete deliberation—agreements, disagreements, and successful objections—and commits the final decision. That decision is then converted into a coordinated operational plan with responsibilities assigned across the society.

Synod's demo takes place in a military war room where five generals reason through a crisis, but the architecture is domain-agnostic. The same decision pipeline can be evaluated on entirely different domains without changing its core mechanism.

How we built it

We designed Synod as a deterministic multi-agent reasoning pipeline.

The agents aren't five personas sharing a prompt. Each is an independent decision procedure. Everything after their reasoning is deterministic code: belief updates, the scoring engine, the chair's commitment, and signed receipts. Models reason; code decides.

Each agent is prompted independently with its own doctrine, producing parallel first-pass judgments. Their arguments are exchanged through structured debate, where an agent may make small, bounded concessions when presented with stronger evidence. Every concession is clamped and recorded on the round's receipt, making every revision explicit, auditable, and reproducible.

Instead of selecting a majority opinion, the entire deliberation is passed to a deterministic chair that makes the final commitment. Every intermediate judgment, debate, concession, and final decision is recorded with signed provenance so the complete reasoning process can be inspected and reproduced.

To evaluate the architecture, we built benchmark harnesses, ablation tests, and hold-out worlds—authored by a rival model (Claude) specifically so we couldn't overfit to them—and replayed the system against 4,277 real human negotiations from Stanford's CraigslistBargain corpus.

Challenges we ran into

The biggest challenge wasn't getting multiple agents to talk—it was preventing them from collapsing into the same answer.

Creating genuinely independent perspectives required carefully separating prompts, preserving disagreement throughout deliberation, and ensuring the chair reasoned from the evidence instead of simply following the majority.

Evaluation was another major challenge. A single demo can always look convincing, so we invested significant effort building reproducible experiments, controlled benchmarks, and real-world replay pipelines to test whether the architecture generalized beyond hand-picked examples.

Accomplishments that we're proud of

We're proud that we built an agent society whose central mechanism is preserving disagreement as evidence until commitment.

Every architectural claim is backed by reproducible experiments, ablation studies, hold-out worlds, and evaluation on real human negotiation data.

Most importantly, Synod behaves like a genuine society: independent reasoning, structured disagreement, deterministic resolution, and coordinated execution.

What we learned

Building multiple agents is easy. Building genuinely independent ones is not.

We also learned that demonstrating an AI system isn't enough. If we're making architectural claims, we need evaluation that isolates why the system works, not just whether it works. Designing those experiments ultimately became as important as building the agents themselves.

What's next for Synod

Our hypothesis is that diversity of reasoning is a computational resource, not a liability.Preserving disagreement gives the chair information a single reasoning trace never has.

Next, we want to test that hypothesis across additional real-world decision domains such as recruitment, customer support, procurement, and medical triage. We also want to make societies configurable so developers can compose their own specialist agents while retaining deterministic, auditable decision-making.