See EDDI work, no setup needed
Follow one business decision from question to audit trail: a board of AI agents deliberates, a human approves, and every step is recorded. About two minutes, nothing to install.
The question on the table
A leadership team has to decide whether to commit to a new supplier for three years. Instead of one answer from one model, three specialist agents look at it, each from its own angle.
The questionShould we sign a 3-year contract with a new logistics supplier?
- LegalReads the contract terms
- FinanceChecks price and volume
- RiskLooks for what can go wrong
The board deliberates
Pick a style below. It is the same question and the same three agents, but a different way of reasoning together. Watch what each style does differently.
The questionShould we sign a 3-year contract with a new logistics supplier?
- LegalReads the contract terms
- FinanceChecks price and volume
- RiskLooks for what can go wrong
Choose how the board reasons
What is differentEveryone answers independently and at the same time. A moderator then merges the answers into one recommendation.
The draft renews automatically for another 3 years unless we give 12 months notice, and it has no exit right for repeated late deliveries. I would not sign it as written.
The rates are 8% below our current carrier, but only if we meet a volume floor. At last year’s volume we stay above it, with a thin margin.
This supplier would carry most of our freight. One failure at their main hub could stop our shipments for days.
Sign only with changes: an exit right after repeated late deliveries, a lower volume floor, and a backup carrier for at least 20% of volume.
Sign, but only with three changes
Recorded dissent: Finance would accept a backup carrier for 15% of volume instead of 20%.
What is differentTwo sides argue for and against, then answer each other. An impartial judge scores the arguments and names a winner.
Sign. The rates are 8% below our current carrier and the capacity guarantee covers our busiest season. Three years locks in that price.
Do not sign as written. Three years with one supplier, no exit right and no backup carrier is a risk we cannot undo.
The saving is certain, and a backup carrier can be added later.
Later is the problem. Once we sign, we lose the leverage to ask for an exit clause.
Con is better supported. The saving is real, but the contract makes it hard to reverse. Renegotiate the exit and backup terms first.
Renegotiate before signing
Verdict: Con wins. Score: Pro 4/10, Con 9/10, judged on argument quality and evidence, not on how forcefully it was put.
What is differentExperts estimate in isolation, see only an anonymous summary, and revise. Rounds repeat until the estimates converge.
Late deliveries per 100 shipments in year one. A: 12. B: 4. C: 7. Nobody sees who said what.
After seeing only the anonymous summary (range 4 to 12, median 7), they revise. A: 9. B: 6. C: 7.
The spread narrowed from 8 points to 3. The panel has converged, so no third round is needed.
Plan for about 7 late deliveries per 100 shipments in year one. The contract needs delay penalties from 5 per 100 and a review after 12 months.
Sign with delay penalties and a 12-month review
Convergence: Two rounds. The spread narrowed from 8 points to 3. Individual estimates stay anonymous between agents.
What is differentOne agent is assigned to attack the emerging consensus. The others must defend their position or change it.
I recommend signing. The rates are lower and the supplier guarantees capacity.
Assume this fails in year two: the supplier loses its main warehouse and we have no backup. What happens to our shipments?
The guarantee does not cover a lost warehouse. You are right. We need a backup carrier for part of the volume.
I can add a right to terminate if delivery targets are missed. That answers the exit problem.
The challenge changed the outcome from “sign” to “sign with conditions”: a termination right and a backup carrier.
Sign with a termination right and a backup carrier
Recorded dissent: Risk still prefers a 12-month pilot over a 3-year term.
What is differentOne agent drafts, the others critique, and the author revises. The recommendation is the revised version, not the first draft.
Sign for 3 years. It saves 8% against the current carrier.
The saving depends on a volume floor you did not mention. If we ship less, we pay for volume we do not use.
There is no backup carrier in the draft.
Revised: sign for 3 years only with a lower volume floor and a backup carrier for 20% of volume. The saving falls to about 5%.
Both critiques are addressed in the revision. One item stays open: Legal still needs to negotiate the exit clause.
Sign the revised terms once the exit clause is agreed
Recorded dissent: Legal: the exit clause is not yet negotiated.
What is differentThe question is split into tasks. Agents work on them, and a different agent verifies each result before it counts.
Task 1: Legal reviews renewal and exit terms. Task 2: Finance calculates the break-even volume. Risk then verifies both.
Done. Automatic renewal and no exit right found. A proposed replacement clause is attached.
Done. The new rates beat the current carrier as long as we ship at least 85% of last year’s volume.
Task 2 checked against last year’s shipment data: correct. Task 1: the proposed clause closes the gap.
Both tasks are verified. Sign with the replacement exit clause and a volume floor at or below 85%.
Sign with the new exit clause and a volume floor of 85% or less
Task status: 2 tasks planned, 2 completed, 2 verified by another agent.
What is differentAgents with competing interests trade concessions until they reach an agreement, instead of one side winning.
I want the 3-year term. It secures the 8% discount.
I want 12 months with an option to renew. No lock-in before the supplier proves reliable.
A 3-year term with a right to terminate after 12 months if on-time delivery falls below target.
I accept the termination right and a backup carrier, if the discount applies from day one.
I accept the 3-year term, if the backup carrier covers at least 20% of volume.
Agreement: 3-year term, termination right after 12 months for missed delivery targets, backup carrier for 20% of volume, discount from day one.
Sign the 3-year term with the agreed trade-offs
Concessions recorded: Finance accepted a termination right. Risk accepted the 3-year term. Both concessions are on record.
A mock conversation to show how each style works. Real boards use your own agents, data and rules.
A human decides in Slack
The board only recommends. The group pauses before it concludes, and a named approver receives a card in Slack. Try the buttons.
Approval needed: Group phase Synthesis
- Question
- Should we sign a 3-year contract with a new logistics supplier?
- Recommendation
- Sign, but only with three changes
Only approvers on the allow-list can decide. Every decision is written to the audit ledger.
- Only people on the approver allow-list can decide. Without the list, the buttons are not shown.
- Approve or reject: the decision is attributed to the person who made it.
- The same approval is also available in the EDDI Manager, through the API and through MCP.
Everything lands in the audit ledger
Each step of the run writes a signed entry: what the agent was asked, what it answered, and who approved. Entries are HMAC-signed, append-only and sequence-checked, so a gap or an edit shows up.
Example entries- Finance agent gives its opinionSignature verified
- Sequence
- 1
- Time
- 2026-03-12 09:14:07 UTC
- Agent
finance-advisor- Task
- LLM task (opinion)
- Signature
HMAC v4:9f2c41ab…e81aEd25519 · a41f…09c2
- Moderator writes the synthesisSignature verified
- Sequence
- 2
- Time
- 2026-03-12 09:14:41 UTC
- Agent
board-moderator- Task
- LLM task (synthesis)
- Signature
HMAC v4:3b07d5c9…14f2Ed25519 · 7be0…3d58
- Approver decides in SlackSignature verified
- Sequence
- 3
- Time
- 2026-03-12 09:19:03 UTC
- Agent
board-moderator- Task
hitl.approval- Decided by
slack:U0EXAMPLE- Signature
HMAC v4:c81e6a20…4d03Ed25519 · 52d9…b6e1
An integrity check over the whole conversation reports INTACT, BROKEN or INCOMPLETE.
Now try it on your own case
The example above is a mock. We can build the same flow on your documents, your systems and your approval rules.
Prefer to click around the real thing? Open the raw EDDI sandbox and explore the Manager yourself.
Open the sandbox ↗Demo system for exploration only. Data is wiped every 48h at 03:00 UTC.