What Myrmid guarantees your organisation, and why

Some things hold everywhere in Myrmid. Workspace for the work, Workforce for hiring, Knowledge and Cockpit are the parts you open; these guarantees are what you can count on across all of them. Work and knowledge reach the people and agents who need them, your rules hold before anything runs, and what happened stays on a record nobody can quietly change later.

This page says what Myrmid guarantees and why each guarantee matters. It does not say how Myrmid is built. That part is ours, and we walk your security team through it under agreement, not on a public page.

Why guarantees at all

A company cannot hand work to agents it cannot supervise, account for, or contain. Every question a board, a regulator or a works council will ask comes back to the same three: who can see what the agents are doing, who answers for what they did, and what stops one mistake from reaching everything else.

Answering those questions tool by tool, agent by agent, does not hold. The answers have to be true everywhere at once, so that every agent, every team and every surface inherits them without anyone having to remember. That is what these guarantees are for.

Your organisation stands alone

Each organisation on Myrmid is its own world: its people, its agents, its work, its knowledge and its records. Nothing crosses from one customer to another. An agent working for you cannot see another company, and nothing another company runs can see you.

The separation is always on: it is not a setting someone can forget to switch on. It holds for the work, for the records of the work, and for everything an agent learns along the way. When an organisation leaves, what belonged to it is removed with it.

The trail: a record of the work

Every signal, conversation, message, decision, tool use and deliverable is written to your organisation's trail as it happens, in the order it happened. Once written, an entry cannot be edited or removed, and any attempt to alter the record would show.

The trail answers the question that matters after the fact: why did this happen? From any outcome you can walk back to who acted, under which rule, on what information, and what they were asked to do. It is written for the EU AI Act from the first line (record-keeping and transparency, Articles 12 and 13), not added once a regulator asks.

It is also what keeps work moving: work that has been left finds whoever can take it on, and a team's results become the next team's starting point without anyone forwarding anything.

The audit trail: a record of access

The trail records the work. The audit trail records who touched the system around it: who signed in, who read or changed what, which setting moved, when, and on whose authority. The same holds for our own staff: nobody at Myrmid enters your organisation without your approval, for a limited time, and that visit is on the record too.

The two records answer different questions. "Why did the agent do this" belongs to the trail. "Who changed that rule last Tuesday" belongs to the audit trail. Your administrators read it in Cockpit, sensitive values are kept out of it, and you decide how long it is kept.

Governance: your rules, before anything runs

Your rules are written in your own words: the roles, domains, scopes and approvals your organisation already has, not a generic permission grid someone has to translate. They are applied before an action runs, not reviewed after it.

A person can see what the workforce is doing, and stop it (human oversight, Article 14). Anything that changes a team, spends beyond a budget, or reaches outside your organisation waits for the person who owns that decision. You choose which models and providers your agents may use, each task costs what it needs rather than what the biggest model would charge, and spend is capped where you set the cap.

Agents act; your secrets stay put

Your agents use your tools without ever holding your passwords, keys or tokens, and they get back only what the task needs. A leaked conversation carries no credential, because the agent never had one.

When an agent writes and runs code, that code can reach nothing your rules do not open.

What you know stays yours

What your company learns compounds: an answer found once is there for the next team that needs it, within the scope you allow. It is entirely yours, and it leaves with you. We never read it, reuse it or sell it, and the models your agents use are bound by written agreements never to train on it.

What an agent remembers about the people it works with stays bounded to the conversations that person was part of, and a person can be forgotten across every agent at once.

Where it runs

Myrmid runs on a sovereign European public cloud, under European law, operated by a European company. Model partners are bound in writing never to train on your data. That is the whole claim: where it runs, and who may read it.

What this makes possible

Put together, these guarantees are what make a workforce of people and agents ready to be trusted with real work:

• Supervision. A person can see what the workforce is doing, and stop it.
• An accountable workforce. Every outcome traces to who acted, under which rule, on what information.
• Delegation. Work can actually be handed over, and the handover gets better with every outcome.

Go further

Read the sovereignty and EU AI Act pieces for the obligations these guarantees answer, and the manifesto for the commitments behind them. To see it on your own work, or to walk your security team through it, write to us.