Production is read-only by default
The tools an agent is given for a live system are read tools. That is enforced by how the tooling is set up, not by good intentions.
Approach
How I use AI on your systems, and how every result is checked
An AI agent is an AI model that is given a task, tools and limits, and works through the task step by step instead of answering one question. Agents make this kind of work much faster. Used carelessly, they also make it easy to be wrong at scale. This is how I get the speed without making the work riskier. The site you are reading was built the same way.
The rules
Six rules I hold to
A person applies every change
The agent prepares the change and the proof. Someone with the authority to do so reads it and runs it.
The work is checked independently
A separate verifier re-checks the result against the source systems on its own, record by record, instead of a person doing spot checks.
Safe to run again
Work is designed so that running it twice changes nothing that is already right. A failed run can simply be run again.
Evidence, not summaries
Each run produces before-and-after records for everything it touched. That is what you show an auditor.
Agreed before it starts
Which tools may see which data is settled with you in writing before any work begins.
What enforces the rules
A rule that depends on an agent behaving well is not a rule. Three layers do the enforcing:
- Access. The tools an agent is given for a live system are read tools, with read-only credentials wherever the system offers them.
- Hooks. Rules built into the tooling stop a listed action, such as a write command aimed at a live system, before it runs. The agent does not get a vote.
- Receipts. Each piece of agent work leaves a record of what was read, what was proposed and what was checked, so a person can review it afterward.
See it run
Two agents and an answer key, on made-up records
The rule that the work is checked independently is easier to believe when you can see it. This is a recording of one agent doing a job and another checking it, with the result scored against plain code. On a real migration this step is ordinary code; an agent does it here only so there is something for the verifier to check.
A recording, not live AI. Recorded on October 10, 2026 from a real run on 43 made-up award records. No real students, schools or amounts.
The job
43 made-up records and eleven written rules for bringing a new system into line with an old one. Some records are missing, some are wrong, some are duplicated, and a few are traps, such as a record that matches an award the old system had canceled.
The worker
An AI agent proposed what to do with every record: leave it, create it, correct it, move it, remove it, or hold it for a person.
The verifier
A second agent was given the rules, the records and the worker’s answers, but not its reasoning. It worked out every record again for itself, then compared.
The answer key
Plain code that neither agent had seen produced the right answer for every record. Both agents were scored against it.
- 43 of 43
- records the worker got right
- 0
- records where the verifier disagreed with the worker
- 43 of 43
- records where the verifier’s own answer was right
- 7
- records correctly held for a person instead of being guessed
On this run the worker got every record right, traps included. The verifier, giving its own answer for each record, agreed with all of them, and the answer key confirmed both. A check that finds nothing has still done its job: the result now rests on two separate passes and an answer key, not on one agent’s word.
Show the technical detail
Worker: Claude Haiku 5.5. Verifier: Claude Opus 5.5. Each agent ran once, with no tools, and returned a structured answer for all 43 records. The verifier received the worker’s claimed actions only. The answer key is a short function that encodes the same eleven rules; the records, the rules, the key and both agents’ raw answers are kept with this site’s source so the run can be repeated. This shows the pattern on a small, clean problem. On real work the records are messier, which is exactly when an independent check earns its place.
Hands on
Try the method yourself
A method is easier to trust once you have pressed the buttons. These are the pieces of this site you can try or watch, gathered in one place. Any records in them are made up, and the ones that run do so in your browser without sending anything anywhere.
On this page
The recorded run
One AI agent does a job on made-up records, a second agent checks it, and plain code holds the answer key.
Case study
Run it twice
Check twelve made-up financial aid records, fix them, then run again and see that the second run has nothing left to change.
Financial aid automation
You approve, then it runs
The tooling proposes, a named person at your school approves, and only approved changes are applied, with the evidence kept.
Case study
Watch the race
Twelve made-up students go through the same two jobs twice: once on a timer, once started by the record being saved.
Student system consulting
Where does your school stand?
Four situations a school can be in. Pick the closest to yours for a few sentences on what it usually needs.
Workflow automation
How healthy is your workflow estate?
Ten yes, no or not sure questions, scored into one of three bands, with the pages that speak to the gaps.
Services
How an engagement runs
Six stages from the first call to the handoff. Pick a stage to see what happens, who decides and when billing starts.
Home page
Start with your role
Pick the kind of reader you are and the three pages that matter most to you come up.
Where the speed comes from
What the agents are good at
Most integration and migration work is careful reading: thousands of records, hundreds of rules, and the patience to compare them all. Agents are good at the work around that reading: studying the sources, writing the code and the checks, running them, and explaining every difference they find. The record-by-record changes themselves are made by ordinary code, which behaves the same way on every run.
On a recent financial aid migration, that meant one command that corrects the financial aid records of an entire student population and can be run again safely, with per-student before-and-after evidence. The engineering was in the design of the checks, not in typing the fixes.
- Studying real systems and data before anything is designed
- Drafting integration code, queries and reports for review
- Running the same checks on every record, on every run
- Writing the documentation as the work happens
What AI does not do here
It does not make decisions about students. It is given read tools for production, not write tools. It does not replace review by someone accountable. And it does not get the last word on whether its own work is correct.
Fair questions
What people ask about AI on their systems
Will an AI agent have access to our production system?
Read access only, and only where you have agreed to it. Changes are prepared and rehearsed in a test environment, then applied to production by a person.
Does student data get sent to an AI company?
Not unless you have agreed to it in writing. That is your decision, made before work starts. Where the answer is no, the agent works on structure, rules, code and counts, and the record-by-record work is done by ordinary code inside your environment. Where the answer is yes, it is under terms that bar the AI provider from training on your data. The specifics are on the security and data handling page.
How do you know the agent got it right?
I do not take its word for it. A separate verifier checks every result against the source systems, and the checks are designed before the work, not after. Anything that cannot be proven goes back to be redone.
What if something goes wrong halfway through?
The work is built to be re-run. A run that stops partway can be started again and will only change what still needs changing.
Is this just you using a chatbot?
No. It is engineered tooling: agents with defined tasks and limits, rules that are enforced automatically, and a record of everything they did. I put the agents to work every day and build the tooling, the rules and the checks they run under. There is more in where AI agents fit in registrar and financial aid work.
Have a system that needs untangling?
Tell me what you run and where it hurts. The first call is 30 minutes, free, and you will leave with a clear next step whether or not we work together.