Industrial Agents Hackathon · Munich 2026 · Track Tacto

Parley.
The AI buyer that negotiates.

Code does the math, AI does the talking: verified facts and equal-cost trades, built on game theory.

Track Tacto Negotiation Arena Team Clooless parley.sorevo.de
ParleySheet 01/10Rev D

Problem

Every price increase starts as an e-mail.

Gießerei Brandlhuber GmbH & Co. KG Practice case P01

Price adjustment from 01 November 2026

Owing to sustained increases in energy, raw material and wage costs, we are forced to adjust our prices for volute casing HN 50-200 by 9.5 %: from EUR 126.53 to EUR 138.55 per piece.

to: Purchasing, Hollerau Pumpenwerke

To answer it well, a buyer checks

  • 3years of orders, invoices and scorecards
  • 33market price indices
  • 447suppliers, each with notes, news and contracts

The supplier checks every argument against the same data. Only solid facts move the price.

Source: supplier letter of 11 Sep 2026, practice case P01; Tacto Negotiation Arena data package (33,000 rows in 24 tables).

ParleySheet 02/10Rev D

What Parley does

Reads. Negotiates. Signs or walks.

  1. 01 · Read

    The e-mail, the quote PDF and three years of history with this supplier.

  2. 02 · Negotiate

    Up to ten e-mails. Every offer becomes one total cost over the contract term.

  3. 03 · Close

    Accept the cheapest offer that keeps every must-have, or walk away when the alternative is cheaper.

Total cost per roundschematic
opening round 1 round 2 round 3 alternative: no deal best possible deal

Score = share of the possible saving it captures, from the opening offer to the best possible deal.

Accept
ParleySheet 03/10Rev D

How it works

Facts. Priced choices. A logged decision.

01 · Code checks

Supplier facts

Each with its source ID. Profiles built offline, checked by code; we keep only what passes our gate.

02 · Code prices

Priced choices

Two or three packages per round at about the same estimated total cost. The supplier’s pick shows what it values; our steps answer its steps.

03 · Code logs

Logged decision

Accept, hold or walk, with the reason. Templates write the e-mail; the model stays off.

Every e-mail

On-time delivery averaged 81.8 % sc:70210@2026Q3. Package 1: €28.58, 14 days net, 35 days lead time.

template code

Game-theory-inspired bargaining: preference probing and bounded reciprocal concessions (agent/policy.py, packages()). Details: backup B1.

ParleySheet 04/10Rev D

Backup · how it works in detail

Code owns every number. The numbers win.

Code

Dossier

Supplier profile, cost floor, and facts with checked data IDs.

Code

Valuation

Prices every combination of terms as total cost, and finds sheets at about the same estimated cost.

Code

Policy

Anchor, concession steps, guards G1–G9, accept or walk away.

Code

Reader

Reads the reply: status, verdicts, signal sentences.

Model · off

Writer

No measurable gain in our paired tests, so the final runs on templates.

What we optimise

Arena score = Copen − CdealCopen − Cbest

Share of the possible saving. C = total cost over the contract horizon. Hard: every must-have kept, never worse than the opening, walk if the alternative is cheaper. Tie: fewer rounds, in 8-minute waves.

C countsprice × volumeuncovered need + riskpayment interestfreight (Incoterm)late lead timedefects − chargebackstoolingindex path + risksupplier margin

Leverspricepayment dayslead timecontract monthsvolumeeffective dateindex clauseppm · chargeback

ParleySheet B1Backup

Saturday practice · v1 · case P05 · RFQ

Knows the supplier. Moves the price.

Source: our Saturday practice session P05, Parley v1, Sat 24 Oct (replay e642aeb72a327558). Round 1 offered €28.58 paid in 14 days or €28.68 in 30 days, both 35-day delivery, similar estimated cost. Figures and packages from code; the final build writes every e-mail from templates. Dossier: Tacto’s data package. Supplier and portrait fictional.

ParleySheet 05/10Rev D

From our own practice runs · case P02

Our zero. Found, fixed, tested.

Practice case P02 · arena score by build

  1. v1.160.000
  2. v1.220.269
  3. v1.300.562
  4. Final build0.626

Strategic supplier: the brief says a healthy supplier counts, not only the lowest price.

Source: our arena practice runs on P02 (our arena practice log): v1.16 029552b2c280199e, v1.22 0885001f42db6d2e, v1.30 da71ec8f9b9db3e1, final build settings ebe0c9d87c1f3cc8; arena cost breakdown of the v1.16 deal; agent test test_our_cost_matches_the_arena_including_the_supplier_health_charge.

ParleySheet 06/10Rev D

Decision log · Saturday practice · v1 · case P05

Why did it write that? It says.

Parley → DralveckRound 1
anchor at 28.80 on the standing structure (20.0% below)

Cites offer:70240-0 idx:AS-14404 sc:70210. Three packages, about the same estimated cost for us: €28.58, €28.68, €28.41.

Parley → DralveckRound 2
move 4% of the gap (the supplier announced an internal step: hold)

He said he had to check internally. Two packages: €28.64, €28.74.

Parley · closeRound 3
every thread has finished; accept the cheapest standing offer C=244,164

Accepts €28.64, cheapest by total cost over 24 months.

Source: decision journal of our Saturday practice session P05, Parley v1, Sat 24 Oct (replay e642aeb72a327558); reasons verbatim, e-mails condensed. Supplier fictional.

ParleySheet 07/10Rev D

Backup · from our own practice run · case P11

Facts checked. Then it moved.

Supplier · Odlewnia KalwiszRound 0

We are happy to present our offer: €47.00 per unit, quality commitment 6,400 ppm.

Parley · Hollerau PurchasingRound 1

Your on-time delivery averaged 70.8 % over the last two quarters sc:70140@2026Q3, plus two price indices that fell. Three packages, each equal cost for us.

Supplier · Odlewnia KalwiszRound 1

We checked the delivery record data you cited; it holds. Our terms: €46.50 per unit, 300 ppm.

Source: Parley v1 practice session, Sat 24 Oct, decision journal and arena replay. E-mails condensed.

ParleySheet 05/08Rev D

Results · practice, Sat 24 – Sun 25 Oct

Measured in practice.

0.792

equal-weight mean over ten practice cases, v1.38 candidate, no model, one session each, no zeros

0.850

on seven cases

0.658

on the three earlier-labelled hold-outs

Model test · v1.990 paired games, full model minus none +0.0001worst-5 % decision time, without / with 0.21 s / 7.43 sfinal: buyer model calls 0

Identical repeats show it reproduces, not that it is independently validated. Source: our arena practice runs (agent/MERGES.md, v1.38 live record); ablation from the v1.9 test, not measured on v1.38 (README.md).

ParleySheet 08/10Rev D

Real suppliers

Built for suppliers it has never met.

Rules it plays by

  1. Round one is the negotiation.
  2. Lead with the cost drivers that fell.
  3. Never push after a final offer.
  4. Walk away when the alternative is cheaper.
ParleySheet 09/10Rev D

Parley

Lowest total cost. A reason for every trade.

  1. Watch it: every decision is logged with its reason; our runs are in the replays.
  2. Try it: demo, technical video and architecture at parley.sorevo.de.
  3. Team Clooless: built in 25 hours in Munich.
QR code for https://parley.sorevo.de

parley.sorevo.de

Demo, videos and replays

ParleySheet 10/10Rev D