We pair your people with agents that learn the work.

The agent stays on, remembers what your team taught it, and improves between runs. The gain shows up in your numbers, and in what your team takes on next.

The pilot is paid to hit a number.

You choose the performance metric: cycle time, exception rate, first-pass yield or hours returned to the team.

We bring a persistent agent into the systems you already run: ERP, CRM, TMS, document stores or internal APIs. It learns the workflow from the people who do it, and your team gives it a name. We only take real projects that touch money, risk, or volume.

How we work
  1. UnderstandThe agent is the last thing we add. First we map auth, data, exceptions and the people the agent will work beside.
  2. BuildThe agent works real cases with your team, and our harness has it review its own misses until it holds the agreed number.
  3. ApproveYou review the result and the access it needs.
  4. DeployIt runs where you said it would, and keeps what it has learned.

Infrastructure built for agentic workflows.

A persistent agent needs a place to run that keeps its memory and tools between sessions. We design that environment around your agents, in your infrastructure or Marbell Cloud.

AI infrastructure

Gate decides each request before any data moves.

Gate is a catalog of approved services with a policy engine in front. SSH hosts, Postgres databases, Kubernetes clusters, MCP tools and models are listed by name rather than by network route. An agent asks for a single resource. Policy allows or denies it before a byte moves.

A request carries an identity, an action, a resource and a context. Policy, written in Cedar, evaluates all four. The audit log records the request and the decision.

  • Grants: a person hands an agent a subset of their own access, such as a single database. The agent never inherits the full account.
  • Revocation: pull the grant and the next check is a deny. There is no password to share or rotate.
  • Lifetime: a grant lasts as long as you set.
  • Identity: OIDC proves who is asking. Gate decides what they may touch. A CI job and an engineer follow the same rules.
How Gate works

We engineer human + AI superintelligence.

By that we mean a person who knows the business, paired with an agent that keeps learning it, doing work neither would manage alone. Our leadership comes from cloud infrastructure and quantitative machine learning, from nearly twenty years of running Swiss companies, and from management consulting, executive coaching and change management. We're based in Rotkreuz, Switzerland.

Tell us about the work, and the people who do it.

Describe your project

A film about a pilot project

The pilot · 1:24 Start
  1. boot marbell-agent --pilot. RSI harness online.
  2. cat AGENTS.md. Persistent agent, Marbell RSI harness.
  3. Discovered: SEIM pilot project. 14 tools found.
  4. Operator: What can you do?
  5. Agent: Let me show you... but first, what is my name?
  6. Operator: Max.
  7. build graph --seim
  8. Map of contents.
  9. Critical security alert. Unusual sign-in burst, host-01.
  10. Acknowledged.
  11. Max requests: checkmk read access?
  12. Grant: checkmk, read.
  13. Operator: We have a Friday deadline, can you help.
  14. Max: On it.
  15. Reports ready, 3 of 3.
  16. Friday 17:00. Done.
  17. Saturday 02:47. Max: I need you at the console, I found something.
  18. Critical: anomalous login burst, edge-02, 02:47.
  19. Blocked, 02:48.
  20. Operator: Thanks, nite.
  21. Max: Nite.
  22. Where do we go from here?
  23. marbell.com/start. Pilot Projects that make a difference.