Case Study: The Integration Agent. How One AI Agent Keeps a University's Systems in Step
From One Hire to a Whole Economy explained the idea with a timesheet check: one AI agent, hired again and again, by anyone who needs that job done. This case study follows a different agent, one that anyone who has worked in a university's IT department will recognise the need for: an agent that keeps the university's systems in step with each other.
The companion to From One Hire to a Whole Economy. One AI agent keeps a university’s systems in step: hired again and again by admissions, student records, finance and a partner university, with a signed record of every change. What it does, what it never does, and how it would run on AWS.
- From fifteen links to one agent. Six systems wired directly to each other can need up to fifteen links. With an integration agent, each system has one way in.
- One change, every system. A student accepts an offer; the agent updates records, the learning platform, finance and the library, and returns one signed receipt.
- Hired again and again. By admissions, student records, finance and a partner university, each hire with its own purpose, limits and receipt.
- Built from services that exist today. One example stack on AWS: EventBridge, Step Functions, Lambda, Bedrock, KMS and S3 Object Lock. Azure and Google Cloud have equivalents.
Bottom line. Every university has the same integration problem. Very few need to solve it on their own.
Below: the full article, for technical teams ↓"Every university has the same integration problem. Very few of them need to solve it on their own."
Start here
This is the companion to From One Hire to a Whole Economy. If you have not read it, the five-minute version is in the explainer. The terms used below, capability, manifest, hire-token, receipt and registry, are defined there.
Example University is invented. The systems it runs are the ones most universities run: a student record system (for example Ellucian Banner or Tribal SITS), a learning platform (for example Moodle or Canvas), an admissions system, finance, the library and the ID card system.
The problem every university knows
When a student accepts an offer, at least five systems need to know. The student record system needs a record. The learning platform needs the right modules. Finance needs an account for fees. The library and the card system need the student too.
In most universities, each of those connections was built separately, over years. Some are direct links, some are files that move overnight, some are a person re-typing data. Every new system means new connections. Every upgrade risks breaking one. When something goes wrong, nobody can easily say which system changed what, and when.
Before and after: from fifteen links to one agent
- A point-to-point link, built and run by the integration team
- The integration agent, hired for one change at a time
- A signed receipt for every change
Six systems wired to each other need fifteen links. With an integration agent, each system has one way in, and every change leaves a signed record.
The integration team spends its time maintaining connections instead of improving them. And the student, on their first day, finds they cannot log in to the learning platform because the overnight file did not run.
Meet the integration agent
Now imagine the connections are not built inside each pair of systems. Instead, there is one capability whose whole job is keeping records in step: an integration agent. Any system that needs a change carried to the others hires it, for that one change.
Its manifest, its CV, is short and strict:
- It does: take one change (a student accepts an offer, changes course, withdraws, transfers), translate it into each system's own language, make the updates through each system's official interfaces, check that every update landed, and sign a record of all of it.
- It never does: change marks or results, create or remove a student without a change in the student record system, decide anything about a student, or keep student data after the job; the receipt records what changed, not a copy of the records.
- When something fails: it does not leave systems half-updated. It either completes every update, or reverses the ones it made and reports the one that failed, in the receipt.
Most of what it does is rules: field mappings, codes, formats. AI earns its place in the gaps: matching a module code that one system writes differently from another, spotting that two records describe the same person, explaining a failure in words a person can act on. Every one of those judgements is shown in the receipt, so a person can check it.
One integration agent, hired again and again
- One integration agent is listed once, with a manifest: what it does, what it never does, and its price.
- Four different hirers use it, each for its own change: a new student, a course change, a fee, a transfer.
- Each hire is its own contract, with its own purpose and limits, and leaves its own signed receipt.
- Every accepted job adds to one track record, which the next hirer, at this university or another, can read.
The same agent, the same manifest, four different jobs. Each with its own contract and its own receipt.
One change, end to end
Here is one hire, in four steps.
One change, in four steps
- 1Hear the changeA system announces that something happened: a student accepted an offer.
event - 2Hire the agentFor this one change, with the purpose, the limits and who it acts for.
hire-token - 3Update every systemTranslated into each system’s language, through its official interface.
official APIs - 4Sign the receiptEvery change listed; if one failed, what was reversed.
receipt
The same four steps for every change, whichever system is hiring.
A student accepts an offer. The admissions system hires the integration agent for that one change. The hire-token says what it is for (create the new student's records), the limits (this student only, these systems only, data stays in the EU), and who the agent acts for (the admissions office).
The agent then fans the change out to every system that needs it:
One change, every system
One hire fans out to every system that needs the change, and comes back as one signed receipt. If any update fails, the receipt says what was reversed.
The admissions system gets back one result and one signed receipt. The receipt lists every system that was updated, what changed in each, which mappings were used, and where AI helped. If the learning platform update failed, the receipt says so, and says what was reversed.
The student, on their first day, can log in. If they cannot, the support desk finds the receipt for that student in seconds, instead of searching five systems' logs.
The same agent, hired again and again
The admissions system is not the only hirer. The same agent is hired:
- By the student record system, when a student changes course, so the learning platform and the timetable follow.
- By finance, when a fee is paid, so the student's access can be confirmed in every system that depends on it.
- By a partner university, when a student transfers, so the right records arrive in the right systems, in the receiving university's own format.
Each hire is separate: its own purpose, its own limits, its own receipt. What the hires share is the track record. Every accepted job, whoever it was for, adds to the evidence the next hirer reads.
Not the only agent
The integration agent is not the only one Example University hires. When an application arrives, the admissions system also hires an admissions document check: an agent that checks the application pack for missing documents and details that do not agree, and never decides; an admissions officer does. It is a smaller agent with sharper limits, which is why Parts 7 to 9 build it first, step by step, before the same pieces are applied to the integration agent.
The three levels, applied
Part 4 looked at the Agentic Economy at three levels. Here is what each one means for the integration agent.
One hire. One change, carried to every system that needs it, with one signed record. The person affected is the student, who gets working accounts on day one, and can be shown what happened to their records if they ask.
A market. Many universities run the same few systems. An integration agent built once, with connectors for the common student record systems and learning platforms, can be hired by many universities, and its track record shows how reliably it keeps systems in step. A supplier that does this well is hired more. The people most affected are the integration teams: their work moves from maintaining connections to setting the rules, reviewing receipts and owning the limits.
An economy. Integration itself becomes something you hire by the job, not a project you fund every time a system changes. The skill that becomes more valuable is the engineering judgement: which systems are the source of truth, what must never be changed automatically, and what happens when two systems disagree.
The EU AI Act question
Much of what this agent does is integration logic, not AI. Where AI is used, it is for matching and explaining, not for decisions about students. The Act's high-risk list does name education as a category, including AI used to decide access or admission, and it allows exceptions for narrow procedural tasks that do not decide the outcome. Whether a given capability falls inside is for each university's legal advisers. What the design can do is make that conversation easy: the manifest states the intended purpose, every use of AI is recorded in the receipt, and the agent never decides anything about a student. Data protection law applies throughout, because student records are personal data, which is why purpose and limits travel in every hire-token.
How it would run on AWS
The table below is one example of a technology stack, to show that this can be built today from services that already exist. It is one way to implement it, not the only way, and it is subject to change as the build goes on and the services evolve. I use AWS in this series; Azure and Google Cloud have their own equivalents for each piece.
| The piece | What it does | On AWS |
|---|---|---|
| The change | A system announces that something happened | Amazon EventBridge |
| The agent | Carries one change to every system, in order, and reverses it if one step fails | AWS Step Functions, running AWS Lambda functions |
| The judgement | Matches codes and records that do not line up, and explains failures | A model on Amazon Bedrock |
| Its limits | Refuses anything its manifest says it never does | Amazon Bedrock Guardrails, plus permissions the code simply does not have |
| Its CV | The manifest: what it does, never does, costs and promises | A versioned document in Amazon S3 |
| How it is called | By other AI agents and by ordinary software | An MCP endpoint through Amazon Bedrock AgentCore Gateway, and Amazon API Gateway |
| The registry | Listings and track records | Amazon DynamoDB |
| The hire-token | Signed permission for one job, checked before any work starts | AWS KMS asymmetric signing keys |
| The receipt | A signed record of every change that nobody can alter | A KMS signature, stored with Amazon S3 Object Lock in compliance mode |
| Measurement | Checks every job against what was promised | Amazon CloudWatch |
Whatever the stack, two design choices matter more than any service name. First, everything stays in one AWS Region inside the EU that has the services you need, and the manifest says which. Second, an integration agent that touches student records can run inside the university's own cloud account, so the data never leaves it; the hire-token and the receipt work the same way wherever the agent runs.
Where my research goes next
These are the questions my next pieces of research will take on. Stay tuned.
- Who owns the truth when systems disagree. The agent can detect a conflict and report it; which system wins is the university's decision. I will set out a simple way to write those rules down, so the agent can follow them.
- How far automatic reversal should go. Undoing a half-finished update is safe for some systems and risky for others. I will map which is which, system by system.
- Whether system suppliers build this themselves. If every supplier ships its own integration agent, universities could end up back where they started. I will look at how a shared standard avoids that.
Coming next
This case study showed what the integration agent does. Next, the series builds an agent, step by step, on AWS, starting with the admissions document check:
- Part 6: On AWS Bedrock, building a capability and its manifest.
- Part 7: On AWS Bedrock, the hire and the signed receipt.
- Part 8: On AWS Bedrock, measure, track record and release.
- Part 9: When an agent hires an agent. Chains of hires across organisations, with purpose and receipts carried all the way down.
- Part 10: Building the marketplace. Listing, finding and paying for agents at scale, and the rules a marketplace itself has to follow.
If you build AI agents and would like to try one against the registry as it takes shape, message me on LinkedIn.
The small print
I am a technical architect, not a lawyer. Example University is invented, and the systems named are examples of what universities commonly run, not a description of any institution. Where I mention the EU AI Act or other rules, I describe the questions they raise, not how they apply to any organisation; this is not legal advice. AWS services and their availability change; check the current documentation before building.
Important notice. This article is part of my ongoing research into the Agentic Economy and reflects my thinking at the time of writing. It is shared in good faith, for information and discussion. It is not legal, regulatory, technical, financial or other professional advice, and it is not a verdict on how the EU AI Act or any other law applies to any organisation, system or situation. Laws, guidance, technology and my own research all move on, and others may reach different conclusions. The examples, designs, diagrams and technology choices show one possible approach. Before relying on, adapting or applying any part of it in your own situation, you must test and verify it, design it for your own architecture, and take independent expert advice, including from qualified lawyers and from your own IT, security and data protection specialists. Any decision to act on it, and its consequences, rest with you. To the fullest extent permitted by law, the author accepts no responsibility or liability for any loss, damage or consequence arising from any use of, or reliance on, this article, which is provided without any warranty of accuracy, completeness or fitness for a particular purpose. Reading it does not create a professional or advisory relationship with the author. The views expressed are the author’s own.