Guardrails

Decide what an agent may see and do

Every agent carries its own guardrails: whether it may change data, which actions wait for a person, and which sensitive values are hidden from the AI model. They are settings on the agent, and they are enforced while it runs.

Pick a case to follow one request through the checks.

Three guardrails, set on each agent

They sit in the agent's set-up, next to its tools. Each one acts at a different moment.

1 · When the agent is put together

Read-only

An agent can be set to read-only. Tools that change data are then taken away before it starts, so there is nothing for it to misuse. A tool that has not been marked as read or write asks for approval every time.

  • How a tool is markedBy its tag. List and Get read. Create, Update and Delete write.
  • What the agent seesOnly the tools it is left with. A removed tool is not offered to the model at all.
2 · Just before a tool runs

Require approval

Name the tools that must wait for a person. In chat, the agent stops and shows what it wants to do. You approve it, edit it or reject it. A request nobody answers counts as rejected.

  • In a scheduled runNobody is there to ask. Those tools are skipped, unless an administrator approved them by name when the schedule was saved.
  • In multi-agent workThe lead agent also asks before it starts a step that takes an action.
3 · On the way to the AI model

Sensitive data

Choose which kinds of value must not reach the model, and what to do with each one. The rule is applied to what the user typed, to earlier replies and to the results that tools bring back, before the model reads them.

  • Built inEmail addresses, credit card numbers, IP addresses and web addresses.
  • Four ways to treat a valueMask it, remove it, replace it with a hash, or block the step.
  • Your own patternsAdd a rule for anything else: a national ID, a phone number, a postal address, an internal account code.
Built for your rules

Your data has its own sensitive values. We build the rules for them with you

Every company has identifiers that no general list covers: a patient number, a loyalty card, a supplier contract code. A custom rule describes the pattern, and you choose how the value is treated. Smart IS works with your team to write and test those rules.

  • Patient numberPT-204-88163mask
  • National IDa pattern for your countryblock
  • Phone number+00 000 000 0000remove
  • Contract codeCN/2026/00417hash

Limits that are always on

These are not settings to remember. They apply to every agent.

A step limitA run stops after a set number of steps, 30 unless you change it.
Only its own toolsAn agent can call the tools it was given, and nothing else.
Checked argumentsEvery tool call is checked for types, allowed values and required fields.
Safe follow-up queriesQueries on saved results are read-only and time-limited.
On record

What is kept, so you can look back

A guardrail is only useful if you can see what happened afterwards.

  • Every chatWhich agent and which connections answered, and each tool call with its name, its arguments with secrets hidden, whether it worked and how long it took.
  • Every scheduled runThe run, its steps and each call, with the tool, the connection, the environment, the row count and the status.
  • Multi-agent workThe list of steps, including any that a person declined.
  • Every change to an agentThe git history and the pull request that approved it.
Good to know
  • Guardrails are set per agent. There is no single switch for the whole organisation today.
  • Sensitive-data rules hide values from the AI model. Saved result files and the tool panel still hold the original data, for the people allowed to see it.
  • The rules apply inside that agent's own steps. When it works under a lead agent, the lead agent and the report agent read the saved data as it is.
  • For work where no data at all should reach a model, use Operational Mode or bring your own model.