The agent reads and drafts in the tools your team already uses, such as Checkmk, your SIEM or the ERP, and it never borrows a staff login to do it. Where you give it write access, every change it makes is logged with the input that led to it, so you can always trace why it acted.
Example: an alert from Checkmk or a SIEM
Take a Checkmk alert. The agent gathers the evidence and drafts the incident note, and your engineer decides what happens next. A draft has no effect until someone on your staff confirms it, which is why we start with draft access. Once the drafts hold up, you can give the agent write access for the steps you trust, while the rest keep waiting for a person.
The same four steps work when the trigger is an order in the ERP or a ticket in the CRM.
An illustration of what an agent with these access levels could do. It does not describe work for a customer.
01 / 04
Read
A Checkmk service alert or a SIEM detection comes in, and the agent pulls out the host name and the failed check.
02 / 04
Read
It looks at the host's recent metrics in Grafana and the related events in the SIEM, the same places your engineer would look first.
03 / 04
Read
It flags what the record is missing, such as a host with no owner on file, so nobody has to find that out at 2 a.m.
04 / 04
Draft
It writes the note with the evidence attached, and an engineer confirms it. With write access it could also acknowledge the alert in Checkmk.
If you run Gate, the agent can also reach the other resources Gate protects, from databases and Kubernetes clusters to the tool servers (MCP) that AI assistants call.
How the result is measured
We test every change to the agent against 20 to 50 of your past cases before it touches live work. During the pilot we also track how much of the workflow the agent finishes and how much checking its output still needs, because the second number decides whether it saves your team time.
Engagement stages
Pilot
Each further system needs about 3 hours of your team's time
You choose read, draft or write for each system separately, so the agent can read your SIEM freely while every change in the ERP still waits for a person. Where Gate is used, the agent reaches each system through a local address that Gate provides, and Gate's Connector holds the system's credential instead of the agent. Nothing on the agent's host could be reused to get around Gate. If you revoke the agent's access in Gate, its very next request is refused, without waiting for a session to expire.
Access levels the agent can hold on a connected system, with examples
01 / 03
Access levelWhat the agent doesWho confirms
Read Retrieves records, for example metrics from Grafana. None
02 / 03
Access levelWhat the agent doesWho confirms
Draft Prepares a change in the target system, for example an incident note. A member of your staff, before the change takes effect
03 / 03
Access levelWhat the agent doesWho confirms
Write Makes the change directly, for example acknowledging an alert in Checkmk. No confirmation for each change. Each write is logged with the input that led to it.
Gate refuses each routed request unless a written rule allows it, so a forgotten rule fails safe. Rules are written in Cedar, a policy language, and can use inputs like these.
What an audit record names
The Gate page explains how agents get their access, how model calls are decided, and which routes Gate cannot see.
An identity for the agent, time with a specialist for each system, and sample records from the workflow.
By default we call hosted models with zero data retention, so the provider keeps no copy of your prompts. That holds for every vendor we might propose, not just the first, because we only use business APIs that offer zero data retention. If you'd rather prompts never leave the building, the agent can run an open-weight model on your own servers instead. Where model calls go through Gate, its Connector holds the provider key and the agent never sees it.
Where processing can run
Your own network, Marbell Cloud in Switzerland, or a model vendor under zero data retention.
The agent and an open-weight model run on servers you control, so prompts stay inside your network.
Gate's Connector inside your network accepts no inbound connections.
The agent runs on our own servers in Switzerland.
Prompts go to the vendor's model under zero data retention.
In Gate, a rule that forbids a model overrides a rule that permits its service.
During Define we run your sample records through several candidate models and compare their accuracy and cost per task. We then propose the model that passes the acceptance test at the lowest running cost, because you pay that cost every month the agent works.
What we sign before discovery, who at Marbell sees your records, and what happens to them if you stop.
For 30 days after release, the engineers who built the agent handle its incidents themselves. They know its quirks, and the first weeks on live work are when quirks show up. After that it's your call: we run the agent under a monthly support agreement, or your team takes it over with the handover pack.
Response times and support hours, the handover pack, and how exceptions, recovery and the licence work.
01 / 07
No. It works under its own identity, so you can always tell its actions apart from any person's.
Through Gate, a person can also delegate part of their permissions to it without handing over a credential. After a revoke or expiry, the agent's next request is denied.
02 / 07
Yes. The agent and an open-weight model can run entirely on infrastructure you control.
If you use Gate as well, its Connector sits beside your systems and accepts no inbound connections, so you don't have to open a port for it.
03 / 07
Not when the agent runs inside your own network. It still has its own identity and an access level for each system. What Gate adds is a policy decision on every routed request and an audit log that shows whether any record was altered.
If the agent runs on Marbell Cloud and needs your internal systems, the placement section on the Cloud page explains how it reaches them.
04 / 07
We test hosted models from OpenAI, Anthropic and Google against open-weight models such as Llama and Mistral, and propose whichever passes your acceptance test at the lowest running cost.
With Gate, you can also permit or forbid individual models, and each call is recorded with the model and its token counts.
05 / 07
No. Your records and prompts are not used to train models for other customers.
Hosted model calls run with zero data retention, so the provider has nothing to train on either, and a custom model trained on your records is used for your project alone.
06 / 07
Yes. For the first 30 days, the engineers who built it handle its incidents, and after that we can run it under a monthly support agreement.
If you'd rather run it yourselves, your team takes it over with the handover pack.
07 / 07
We return your sample records and delete our copies within 30 days. You keep the evaluation report and the workflow map, and your team removes the agent's identities from your systems.
A few sentences on how the work runs today and which systems it touches are enough to start. Your first call is with an engineer, so bring the technical questions.