Work modes
What an agent may do at a customer is not a pile of loose checkboxes. It is one chosen work mode on the task. Whoever creates the task chooses the mode. The platform enforces it on every tool call.
The modes
| Mode | In one sentence | What the source needs for it |
|---|---|---|
| Look only | Read metadata, counts and history. Proves nothing, touches nothing. | A read-only login with permission to read metadata and statistics. |
| Measure on the source | On top of that, the estimated execution plan and the real execution counters, without running anything. | Extra permission to show plans. |
| Prove on a disposable copy | Read on the source, and run and measure the proposal in a copy. | A copy of (part of) the data, with the content the customer allows. |
A mode is a combination of four questions:
- Where is the work done?
- On what: the source itself or a disposable copy?
- What is in the copy: synthetic data, synthetic data shaped to the real statistics, or real data?
- What is allowed: read only, read plus measure, free SQL on the copy, or write on the source?
No two modes give the same answer to those four questions.
Why modes?
A work mode makes visible what a task costs and what it risks. "Look only" needs a read-only login and cannot break anything. "Measure on the source" needs extra permissions, and those extra permissions are exactly why it is a separate mode. "Prove on a disposable copy" lets an agent really run its proposal, but never on the real database.
- A question about structure, quality or history: look only.
- A question about why a query is slow: measure on the source.
- A proposal for a new index, a rewritten procedure or a new view: prove on a disposable copy.
Who checks what
- On creation the platform refuses a mode that does not exist, or a mode the source does not support. The refusal comes before anything runs that costs money.
- On every tool call the tool gateway checks what the tool needs from the mode. A measuring tool in a task on "look only" is refused. The refusal names the task's mode, not the other tools that would have been allowed.
- On screen all modes are listed with one sentence of explanation.
Monitors only look
A monitor always runs in "look only". A periodic check reads and judges, but proves and changes nothing.
And writing on the source?
Writing on the real source is not a work mode but a permission per source. An agent can propose a script. A human approves it, and only then is it run. Without that permission on the source nothing happens, not even after approval.
Example
A BI developer wants to know whether a new index makes the nightly load faster.
- She chooses prove on a disposable copy.
- The agent reads the structure and statistics on the source, and creates a copy.
- It runs the procedure in the copy, measures the run time, adds the index and measures again.
- The platform compares both measurements and gives a verdict.
- The developer sees the result and decides herself whether the index goes to production.
More about the copy: The verification sandbox.