Running AI agents in production?
We're working with a small number of teams deploying autonomous agents against sensitive systems. A design partnership is a founder-led engagement: we work on your estate, in your constraints, and what you run into shapes what we build next.
Tell us what your agents touch and what is blocking the security review. That is the whole application.
What we do together
Six pieces of work, in roughly this order. Where you already have one covered, we skip it.
Discover your AI estate
Get the list of agents and MCP servers actually running — including the ones that never went through a review — and find the agents nobody owns.
Map agent permissions
Work out what each agent can reach today: the servers it is connected to, the tools on them, and the permissions that were granted once and never revisited.
Secure MCP access
Give each server its own identity, authorize each connection at the tool level, and inventory the tools so a definition that changes after approval is visible.
Implement runtime controls
Write the rules that matter for your business — including conditions on the arguments a call carries — in observe mode first, then enforce once the decisions read right.
Build audit evidence
Turn the decisions into something you can hand over: an append-only, hash-chained record and an export you or your auditor can verify independently.
Work directly with the Praesidia engineering team
No account-management layer between you and the people writing the code. What you hit gets triaged by the engineers who can fix it, and your constraints shape the roadmap.
Who this is for
The partnership only works when there is something real to secure and someone accountable for securing it.
Your agents are in production
They act against systems that matter — customer data, money movement, internal tooling with real permissions — rather than sitting in a prototype nobody has connected to anything yet.
Agent security has a named owner
Someone on your side owns this problem and can make decisions about it. Engagements without that person stall, and we would rather say so up front.
There is a deadline behind it
An audit, a customer security review, or a regulatory requirement with a date on it. Urgency is what turns a pilot into something both sides finish.
Not there yet? Start on the free plan with a single agent, and come back when the estate is bigger.
What a partnership costs, and what it does not include
Scoped per engagement
Design partner pilots are paid and scoped to the work, with the range published openly on the pricing page rather than quoted differently to each team. Plan limits are set for the pilot, not by a checkout page.
What we are not promising
We are not selling a certification, and framework mappings support your review without constituting one. The controls you get are the ones already built — the trust center lists what exists today and what we do not hold.
What to put in the email
Four lines is enough. We reply with either a call or an honest "not yet, and here is why".
Or start on your own, today
The free plan puts one agent under control without talking to anyone here.