What Agents Can Do
An AI agent connected through the MCP server can do most of what you do in the Rule Repository except the steps that change what runs in production. This page lists what an agent can and cannot do, the limits that apply to it, and what information about your business it can read, so you know what to ask for and what to expect.
Everything below also depends on your own permissions: an agent acting for you can do only what you are allowed to do. If your application offers only the Rule Repository, the evaluation abilities are missing; if it offers only Audit, the agent can only work with evaluations.
Finding and Reading
An agent can:
- List your source objects and their versions, and look up fields, methods, actions and lists of values, by their names, their descriptions, or other words your developers gave them for AI agents.
- Look up the values of a list that comes from your application, such as customers or products, when your application allows it.
- Find rules by name, description, source, whether they are live, or whether they wait for approval, and read any rule or rule version, shown as plain text.
- Give an overview of every rule of a source object: its type, its live version, versions waiting for approval, drafts, when it last ran, and which other rules use it.
- Tell the history of a rule: who wrote each version, when, who approved it and made it live, and what changed from one version to the next.
- Find where a field is used, or which fields no rule uses.
- Read the activity log, when you have permission to see it.
An agent reads the short description of every object, but never the internal notes.
Writing
An agent can:
- Check a rule against a source version before saving it: whether it is complete, unfinished, or wrong, and where.
- Save a rule as a new rule, a new version of an existing rule, or a draft. It asks you which you want before saving, and asks you to confirm before it changes an existing rule.
- Show the impact of a change before saving it: which other rules use the rule, which tests cover it, and how many recent decisions would have come out differently.
- Update drafts, turn a draft into a rule or a version, and change the names and descriptions of sources, rules and versions.
Every rule or version an agent saves waits for a person's approval. It can never make it live.
Testing
An agent can:
- Try a rule with values you give it, such as "Would this rule approve a customer aged 17?", and say which condition decided the answer.
- Write tests from values, from a value template, or from a recorded evaluation, and change them.
- Run saved tests, a few at a time, and report each outcome. The outcomes are recorded on the tests as when you run them.
- Save and change value templates.
When a test would call one of your application's real methods rather than a test method, the agent must have your confirmation first. Your AI client asks you, or the agent asks you in the conversation; each confirmation is good for that run only.
Working With Evaluations
When your application records evaluations with Audit, an agent can:
- Find evaluations by rule, version, source, correlation ID, status, result, person or period.
- Explain an outcome: which rule decided it for a business case, through which branch, and with which values, such as "Why was application 4521 declined?"
- Count and compare: how often each rule ran and what it decided over a period, the rules that never ran, and the share of cases that got a given outcome, by group, or before and after a version went live.
- Replay an evaluation step by step, the way the Evaluation dialog does.
What Agents Can Never Do
Whatever the permissions, an agent can never:
- make a rule version or a source version live, or take one offline;
- request, give or decline approval;
- delete anything;
- read or write internal notes.
These steps are always a person's, in the Rule Repository. See Reviewing Agent Work.
Limits
Your application sets limits on how many tests an agent may run: by default 10 in one request, 100 per hour and 2 at the same time. An agent over a limit is told when it may try again. Your application can also stop agents from reading the values of lists that come from your application, or allow only some of them.
Connecting an agent shares your source objects, rules and counts with the AI client you use. Beyond that, some parts can hold real business data; each is controlled separately:
| Where |
What it can hold |
Controlled by |
| Values of lists that come from your application |
Real records, such as customers or accounts |
Your application's agent settings |
| Evaluations, and tests made from them |
Real production values |
The permission to see evaluations, and to save them as tests |
| Test values and value templates |
Made-up values, or values copied from evaluations |
The permission to see tests |
| Test runs that call real methods |
What your application's methods return |
The permission to run real calls, and your confirmation |
| Values written in rules |
A value or a list item a rule names |
Not controlled separately |