How It Works

A OpenSynod session takes your team from a strategic question to a clear, documented recommendation in a structured, seven-screen flow. Here is what that looks like from start to finish.

The Flow at a Glance

The entry point. You see your team's discussion history — past sessions with their outcomes, in-progress sessions others have started, and saved templates.

Click Start New Discussion to begin.

Step 2: Topic Setup

Define what is being decided.

Decision question — the specific question being put to the panel (required)
Desired outcome type — are you looking for a recommendation, an exploration, or a risk assessment?
Success criteria — what would a good outcome look like? (optional)

Once you select a panel in the next step, a sidebar shows the estimated cost, number of turns, and expected duration.

Step 3: Panel Selection

Choose the expert panel for your decision type. Examples: Go-to-Market Strategy, M&A Due Diligence, Technical Architecture, Pricing Decision...etc

Panels are pre-built configurations optimized for specific decision categories. Each one defines:

Which agents sit at the table (3–7 seats)
Which LLM backs each agent
What persona and domain focus each agent has
Which discussion rules apply:
- Devil's advocate requirement — at least one agent seat is explicitly configured as a skeptic, required to challenge the emerging consensus
- Hidden-position commitment — each agent privately commits an initial position before seeing what others think, preventing the first speaker from anchoring the group
- Source citation requirement — agents must cite sources when making factual claims during the debate phase

Step 4: Discussion Rules

Set session parameters. All fields have sensible defaults.

Setting	Default	What it controls
Speaking order	Round Robin	Order in which agents speak — Round Robin, Dynamic, or Moderator Assigned
Allow human interventions	On	Whether you can send messages during the discussion
Require citations	Off	Agents must cite sources for factual claims
Anonymize agents	Off	Hide agent names and models during the discussion
Max turns per phase	4	How many turns each phase runs
Opening statement words	200	Word limit for opening statements
Rebuttal words	150	Word limit for rebuttal messages

Click Start Discussion to launch.

Step 5: The Live Discussion

This is the core experience. The screen has three zones:

Round-Table Visualization

Agent seats arranged in a circle. The active speaker is highlighted with an animated pulse ring while thinking and generating a response. Each seat shows the agent's persona name.

Human participants are shown separately around the table.

Live Transcript

Messages stream token-by-token as agents think. Each message shows the agent's persona name, model, phase, timestamp, and message text.

Sources cited during the discussion are collected in a Sources tab in the side panel, showing the title, URL, and domain of each reference.

Control Bar

A persistent input bar lets you intervene at any time — type a message to inject a question, request a clarification, or redirect the discussion.

Additional controls let you pause, skip a turn, or end the discussion.

The discussion moves through five phases, managed by a Moderator agent:

Opening positions — each agent states their initial view
Exploration — agents question and build on each other's positions
Debate — explicit challenges, source citations introduced
Convergence — the Moderator synthesizes areas of agreement and disagreement
Vote — the Moderator presents its recommended conclusion

Step 6: Voting Phase

When the Moderator concludes the discussion, the screen transitions to voting.

You see:

The proposed recommendation (or "no consensus" statement)
Key supporting arguments
Substantive dissents — minority positions that have real weight, with their reasoning preserved
Agent votes — each agent's Yes/No/Abstain with a one-sentence rationale

Agents vote first. You see the agent distribution before casting your own vote. Human votes include an optional rationale field.

When voting closes, the final tally is recorded: agent votes and human votes separately. If humans override the agent recommendation, that divergence is explicitly noted.

Step 7: Outcome & Audit Record

The final document. Every session concludes with a structured record that includes:

The recommendation (or "no consensus" with documented reasons)
Vote breakdown — agents and humans separately
Key supporting arguments
Substantive dissent — preserved minority positions
Confidence indicator — how strongly was consensus reached?
Source density indicator — how grounded in cited sources was the discussion?
Full source bibliography
Complete transcript (filterable by agent or phase)
Agent configurations
Participating humans and their votes
Export options: PDF, Markdown, JSON

Teams can also mark the decision outcome later — did we adopt this recommendation? Did it work?