The agent works. Nobody knows what data it reaches, or with whose permissions.
AI agents entered organizations fast, and MCP servers came with them. This programme puts both on the same table: it teaches a deterministic approach to agent design, then inventories the MCP servers already installed and audits authentication, permission scope and data access. What you leave with is an enterprise gateway, narrowed permissions, and a telemetry design that can answer the question after an incident.
Agent and MCP Design Training and Health Check — Training and Health Check
The problem underneath.
Building an AI agent is easy now. What is hard is explaining what that same agent is doing six months later. In most organizations agents were built team by team, MCP servers migrated from developer laptops into the corporate environment, and nobody can see from one place which agent connects to which data source under which identity. That is a visibility gap before it is a security gap.
The programme has two halves, and the order matters. Design first: where an agent decides, where it follows a rule, and which step has to be repeatable. Determinism here is not a preference but a precondition for auditability — an audit log over a system that cannot return the same output for the same input does not say much. Then the scan: the inventory of installed MCP servers, the identity each one connects with, how wide the token scopes are, which source it can read and which it can write.
It runs three days a month for a straightforward reason: this surface does not stand still. A new MCP server gets installed, a team writes another agent, a model version changes. A one-off scan is a photograph of the day it was taken; a monthly rhythm keeps the organization’s agent surface under control over time.
For a public snapshot of where the ecosystem stands today, see our MCP and agent tooling radar; the scan is that same exercise run against your own inventory.
What the programme covers.
Deterministic Agent Design
Where the agent decides and where it follows a rule; state machine, fallback paths and repeatability. The architecture of an auditable agent gets built on a whiteboard during the training sessions.
Inventory of Installed MCP Servers
Which MCP servers are running today, who installed them, which environment they sit in and which data source they touch. In most organizations the list itself is the first surprise.
Authentication Review
The identity each server connects with, where tokens are held, which accounts are shared and which keys never expire; separating access that should be tied to a person.
Permission Scope and Data Access Review
Which agent reaches which source, and on what grounds; whether read and write are separated, and whether scope is genuinely limited to what is needed. Every excess permission is recorded individually.
Enterprise Gateway and Least Privilege
Agents connecting through a central gateway rather than directly; least privilege applied, and the actions that require human approval defined explicitly.
Telemetry Design
Which calls are recorded, how long they are kept, which fields are masked. One question settles the standard: after an incident, can these records answer what happened?
Who this is for — and who it isn’t.
Most vendors only answer the first half. The second half saves everybody a quarter.
A good fit if
- Several teams have each built their own AI agent and nobody sees the whole picture
- MCP servers are installed, but there is no central record of which one reaches which data
- Your agents are going into production and you need to be able to answer an audit question
- You want your team to move agent design from “it works” to “it can be audited”
Probably not if
- You have no agents or MCP installation yet — the Python agent workshop is the better starting point
- What you need is adversarial testing of the model layer rather than the agent surface — AI Red Teaming covers that
- You want a one-off report; this programme produces its value through a monthly rhythm
Questions we get asked.
Is this training or an audit?
Both, deliberately in that order. The audit shows what is wrong; the training decides why the next one will not be built wrong the same way. If you commission only the audit, the same findings come back six months later.
What does a “deterministic agent” mean — isn’t the model probabilistic?
The model is probabilistic; the agent architecture does not have to be. The distinction is this: reasoning left to the model is separated from flow committed to code. Deciding which step has to be deterministic is the real subject of this programme — an audit log only means something over a repeatable flow.
We already have a list of our MCP servers. Is this still worth it?
Having the list does not mean the scopes have been reviewed. In practice the most common finding is not a missing server but an excess permission: a connection granted write access where read would do, or several agents connecting under a single shared identity.
Do we have to stop production?
No. The scan works from inventory, configuration and access records; it does not touch live systems. Changes such as narrowing permissions are written into the report, and you set the schedule for applying them.
Is this the same as AI Red Teaming?
No, but they complete each other. Red Teaming attacks the model and the application: prompt injection, data leakage, jailbreak. This programme looks at the surface around the agent: identity, permission scope, data access and telemetry. An organization needs both.
Often taken alongside.
AI Red Teaming & Security
A 3-day security health check for your AI systems — prompt injection, data leak and jailbreak testing, review of production automations, telemetry audit and a detailed security report.
AI Agent Coding with Python
A hands-on 2-day workshop taking developers from Python basics to coding LLM-based agents with local models, Claude and OpenAI, using LangChain, LlamaIndex and Jupyter Notebook.
Fractional AI Officer (AIaaS)
A monthly-retainer fractional AI officer: three days a month of advisory, direction and technological leadership across all AI work in your company, on a shared calendar.
Let’s talk about Agent and MCP Design Training and Health Check.
Start with a free 90-minute assessment. We map where this fits, what it would touch and what it would change — then you decide.