Patrimoine · How it works

Installed in a day. Run 24/7.

Installed in a day, run by ten agents, under a harness that lets nothing through.

/po

01Installation

Installed in a single day. Three actions on your side.

Three actions on your side: create an empty cloud project, apply the module we provide, connect your SSO. The agents and the pilot do the rest. We never hold admin rights.

Day 1

The base is running

Cloud project, signed images, identities, secrets, sandbox. Started at 8:15 am, in production the same evening on the demo.

Weeks 1 to 2

Your first sources

One or two sources that matter, connected, catalogued, with their first certified KPIs.

Weeks 3 to 6

Run in parallel, then switch

Patrimoine runs next to the existing stack; we compare, we measure cost. Then we switch, flow by flow. Costed review at the end.

02The harness

Agents with a clear mission. Every action goes through the harness.

A model only performs when a safe, reliable structure surrounds it. That is the job of Patrimoine’s proprietary harness: it gives each agent its mission, its tools and its limits, and every action passes through seven layers before it touches anything.

The Quality agent wants to quarantine 312 rows: the action crosses all seven layers before being applied.

01Limited toolsAn agent only has the tools of its mission. Never free access to the machine.
02AuthorisationEvery call is checked against the agent’s permissions and autonomy level. Sensitive actions wait for a human signature.
03The sandboxChanges are prepared and tested on an isolated branch, then submitted for review, like code.
04The log and the traceEvery action, every model call, every cost is recorded. Everything can be replayed and read by your teams.
05ContextCatalogue, contracts, human decisions, past incidents: the agent reasons with what the company already knows.
06The model gatewayAgents call a role, never a model: classify, reason or write. The harness picks the cheapest model that holds the role, under your policy.
07The benchDozens of scenarios replayed on every change, scored by code. An agent or a model that regresses does not ship.

03Every agent has its own page

Its tasks, its cost, its trace. Live, 24/7.

Today’s tasks, open PRs, cost against its budget, its 30-day success rate, and its live trace, tool by tool. Six months of activity in one grid.

The Quality agent’s page, in night mode: tasks, PRs, cost, live trace.

Demo version · Maison Verdier, 03:12.

ModelsPolicy: EU strict

Agent	Role	Effective model	Budget / day	Bench
Quality	classify	Mistral Small	€4	97%
Operations	reason	Mistral Large	€6	95%
Insights	write	Mistral Medium	€3	96%
Catalogue	write	Mistral Small	€2	98%

Example from the Maison Verdier demo, a fictional retailer.

04The models

Your agents’ performance is measurable. The models, interchangeable.

Every agent has a dominant role, an effective model, a daily budget and a bench score. The model is changed in one click, for three reasons: to improve precision on a role, to follow a company policy, or to hold a budget. The agents carry on without noticing.

05Cloud costs

Constant optimisation. Best practices applied continuously.

The agents keep tuning the platform: the right cadence for each table, nothing refreshed more often than it is read, nothing kept running that nobody uses. Each proposal is costed and reversible, and you approve or decline it.

Costs agent proposals · this week

Move mart_web to dailyread once a day, refreshed every hour

−€34/monthAcceptDecline

Pause 32 reports nobody opensnot opened for 90 days

−€310/monthAcceptDecline

Slow down at night and on Sundaysreduced cadence from 10 pm to 6 am and on Sundays

−€120/monthAcceptDecline

Proposed saving−€464/month

Example from the Maison Verdier demo, a fictional retailer.

Next pageTeam Who is behind Patrimoine, the incubator, the agent team.