Agent trust architecture
Designing and building agent trust architecture
I design and build the layer that constrains agents in production, on my own or together with your team. Permissions, approvals and logs are enforced by the server, not by the prompt.
- Scope
- The permission, approval and logging layer of an agent, RAG or MCP system, in full or in parts. We decide together what belongs in it after the intro call.
- Effort
- Design and build, scope by agreement. You receive the proposal after the intro call.
- Outcome
- A design your team understands and can take forward, and on request the code for it, with tests and evals that check the boundaries on every release.
Who this is for
Your team builds an agent that works with internal data or writes to other systems, or a RAG system that searches documents with different access rights. Or a review has shown where boundaries are missing, and now they need to be built.
I have built and run these building blocks myself, for example at Kertos, a multi-tenant compliance SaaS. They include a tenant-separated RAG pipeline, OAuth 2.1 for MCP, declared write operations with a plain-text diff and approvals that stay blocked when something fails.
When another format fits better
- You first want to know where your system stands. Then the go-live review is the better start. Afterwards you know which parts of the layer are missing.
- You are looking for a ready-made product. I do not sell a filter platform or any software. The layer is built in your code and belongs to your team.
- You are looking for operations and monitoring. What gets logged and when an alert fires is part of the design. I do not take on day-to-day operations or run a SOC.
- Network and infrastructure. Which hosts an agent may reach is part of the layer and therefore part of my work. Networks and servers themselves are better hardened by a team that specializes in them.
Contents
What I design and build
-
Least privilege and human in the loop
Every tool gets only the permissions it needs. Where the impact is high, the agent waits for a person.
-
Permissions down to document level
Search filters by the permissions of the person asking, including retrieval. When the permissions on a document change, the index follows.
-
Tenant isolation with isolation tests
Each tenant sees only its own data, and a test fails as soon as that no longer holds.
-
Write approvals with a plain-text diff
Every write goes through a declared operation that the server checks. The approver sees what will change.
-
Audit log
A tamper-evident log traces every operation through a correlation ID from the user’s question to the write.
-
Tracing with OpenTelemetry
Every agent action becomes visible as a trace, including the model that carried it out.
-
Sandboxing and secrets management
Code that the agent runs is isolated, and keys are stored where the agent cannot read them.
-
Masking of personal data
Personal data is masked before it reaches the model, and in test data as well.
-
OAuth 2.1 for MCP
MCP servers and external clients sign in with OAuth 2.1, using tokens bound to the server they were issued for.
Process
How we work together
-
Intro call
We clarify which system it is, how far along it is and which parts of the layer are missing.
-
Design
I design the layer for your system and go through the design with your team before anything gets built.
-
Build
I build on my own or together with your team, in your repository and with your tools.
-
Safeguard and hand over
Tests and evals check the boundaries on every release. Your team takes over code and design and keeps maintaining both.
Writeups
Further reading
In these articles I describe patterns that this page is about, each with code or configuration.
Contact
Talk about your system
In the intro call we clarify which parts of the layer your system needs and how we work together. You then receive a proposal.