AI agent permissions and constraints.
UNCODE turns your policy into AI agent permissions and limits, enforced the moment an agent acts.
Every action checked before it runs.
Least-privilege tools
Each agent gets only the tools its job needs.
Data kept to approved models
Sensitive data goes to approved models only.
Inputs and amounts checked
Inputs checked before anything runs.
Risky actions go to a person
High-impact actions go to a person.
Same rules for every agent
Every agent inherits the same controls.
Enforced in code, not prompts
Write a rule once. It applies to every request.
See where each rule applied
A record shows where each rule applied.
Action requestBefore it runs
An agent asks to pay invoice INV-20418
| Check | Detail | Result |
|---|---|---|
| Tool access | Payments tool | Allowed |
| Data exposure | Approved model only | Allowed |
| Parameters | $15,060, limit $10,000 | Over limit |
| Escalation | Finance lead | Sent for review |
Oversight by risk
More risk, more oversight.
Routine work runs on its own. Sensitive actions wait for a person.
- Summarise a policyRuns on its own, inside guardrails. Low risk.
- Draft a client replyChecked before sending. Reputational risk.
- Update a recordNamed approver. Operational risk.
- Release a paymentApprover and a limit: over $10,000 waits. Regulatory risk.
- Who can reach what?People and agents reach only what their job needs, by team and by role.
- Do our access rules stay intact?Yes. Connectors respect the permissions you already set.
Access by role
| Finance system | Client records | Payments tool | Approved models | |
|---|---|---|---|---|
| Finance team | None | |||
| Care team | None | None | ||
| Finance agent | None | Approval | ||
| Intake agent | None | None |
Write a rule once. Enforce it everywhere.
Agent constraints are limits on the tools, data and actions an AI agent may use, set by risk and enforced in code where the agent acts. UNCODE, an Adelaide AI consultancy, turns your policy into those limits once, then applies them to every agent and request. Rules sit where the agent acts, so a clever prompt cannot talk its way past them. Payments over $10,000 need a person.
Questions about agent constraints.
What are agent constraints?
Limits on the tools, data and actions an AI agent may use, set by risk and enforced where the agent acts, not in a prompt.
Can a clever prompt get an agent past its limits?
No. Rules are enforced in code where the agent acts, not written into the prompt, so a prompt cannot talk its way past them.
Which actions wait for a person?
You decide. Routine work such as summarising a policy runs on its own, while high-impact actions such as record changes or large payments wait for a named approver.
Do agents respect our existing access rules?
Yes. Connectors respect the permissions you already set, and each team and agent reaches only what its job needs.
Set limits before agents act.
Tell us what your agents will do and we will map the limits each one needs.