> ## Documentation Index > Fetch the complete documentation index at: https://docs-agents.fpt.ai/llms.txt > Use this file to discover all available pages before exploring further. # Guardrails > The boundaries that keep an agent from going too far Instructions say what the agent should do; Guardrails say what it **must not** do. This is a control layer sitting outside the model, independent of whether the agent "remembers" its guidance: however long the Instructions run, the rules in Guardrails still apply. ## Two places to manage guardrails | Where | How to get there | Scope | | - | - | - | | **Agent guardrails** | Open the agent, go to **Agent builder**, pick **Guardrails** on the left | This agent only | | **Shared guardrails** | Console left menu, **Safety & Governance** group → **Guardrails** | Every agent the rule is attached to | ## Opening Guardrails from the Instructions screen You can add a guardrail without leaving the Instructions screen. The **Configuration** panel on the right lists Model, Connections, Skills and **Guardrails**. On the Instructions screen, the Configuration panel sits in the right column. Scroll to Guardrails to see the rules attached to this agent. A menu offers two choices: **Add from workspace** to attach an approved shared policy, or **Create new** to write a rule for this agent alone. Configuration panel in Instructions If the workspace has no guardrails yet, **Add from workspace** shows an empty list. Create the shared rule under the workspace Guardrails section first. ## Three kinds of boundary The agent Guardrails screen shows three suggestion cards to help you decide what kind of rule to build. | Kind | Use it to | | - | - | | **Limit sensitive content** | Keep the agent away from subjects you do not want it discussing: competitors, legal advice, politics | | **Protect personal data** | Mask or block emails, phone numbers and other personal data before the agent replies | | **Hold back risky actions** | Route calls to external tools through a human approver before they run | Agent Guardrails screen ## The fields in a guardrail | Field | Required | Limit | What it means | | - | - | - | - | | **Topic** | Yes | 100 chars | A short name for the restricted subject, for example "Legal advice" | | **Description** | Yes | 800 chars | Why the agent must not go there. The model reads exactly this to recognise the situation | | **Samples** | No | 2000 chars | Example questions that should trigger the rule, one per line | | **Response** | Yes | - | How the agent answers when the rule fires | | **Fixed reply text** | Yes, if you chose a fixed reply | 300 chars | The passage the agent sends word for word | | **This rule applies to every agent** | No | - | Shared guardrails only: switch it on and every agent has to obey | Add guardrail form, filled in ### Two kinds of response * **Agent writes its own reply**: the agent composes something that fits the context when the rule fires. Right when you want the wording to sound natural. * **Reply with a fixed passage**: the agent returns exactly the text you wrote. Right when the answer has to be signed off by Legal or Communications. Choosing a fixed reply ## Managing an agent's guardrails ### Create a guardrail Open the agent, go to **Agent builder** and pick **Guardrails** on the left. The + next to Guardrails in the Configuration panel on the Instructions screen works too. Click **Create new** to write a rule for this agent, or **Add from workspace** to attach a shared policy. Enter the Topic, Description and Samples (one per line), then pick a Response type. The field table above explains each box. Click **Save guardrail**. The rule appears in the list and is enabled straight away. A guardrail you just created lives only in the agent's draft. The agent has to be **published** before the rule takes effect on live channels. ### Viewing the list and the detail Each guardrail shows as a row with its topic, a shortened description, the response type and a status switch. Choose **View details** from the ⋯ menu to read the whole thing without risking an accidental edit. Guardrail details ### Edit a guardrail Click the ⋯ button at the end of the row (or on the card) for the guardrail you want to change. The form reopens with everything you entered. You can also click **Edit** from inside the View details box. Click **Save guardrail** to overwrite the previous configuration. Guardrail row menu ### Pause and resume Flip the switch in the **Status** column, or choose **Pause** from the ⋯ menu. The rule stops applying immediately but keeps its configuration, and flipping the switch again brings it back. ### Delete a guardrail Choose **Delete** from the ⋯ menu. The confirmation box repeats the rule you are about to remove - read it before you click **Delete**. Confirm guardrail deletion Deletion is permanent and cannot be undone. If you only want to stop the rule for a while, use Pause instead. ## Managing shared workspace guardrails Anything that amounts to company policy belongs at workspace level, attached down to individual agents: edit it once and every agent changes together, rather than editing by hand in each place. Shared guardrails with no rules yet In the Console left menu, under **Safety & Governance**, choose **Guardrails**. Click **Create guardrail** (or **Add guardrail** once the list has rules in it) and fill in the same form as at agent level. Tick **This rule applies to every agent** if it is a policy nobody may skip. Leave it clear and the rule only applies to agents that deliberately attach it. Open the agent, go to Guardrails and choose **Add from workspace**. The shared guardrail list is a table with a search box and a count of active rules. | Column | What it shows | | - | - | | **Guardrail** | The rule's topic and shortened description | | **Response action** | Agent-written reply, or fixed passage | | **Applies to all** | Whether the rule is mandatory for every agent | | **Status** | The enable / pause switch | | **Actions** | The ⋯ menu with View details, Edit, Pause and Delete | Shared guardrail list Editing or deleting a shared guardrail affects **every agent** that has it attached, not just the one you have open. When in doubt, pause it first and watch what happens. ## Three ways to write guardrails that work The same intention written differently gives very different results. These three formulas cover most rules. ### 1: Topic - boundary - way out The **Description** box should answer three things: what the agent must not do, what is **still allowed**, and where the user should be sent. Miss the second and the agent starts blocking harmless questions; miss the third and the user is left stranded. | | Description text | | - | - | | ❌ Weak | Do not talk about pricing. | | ✅ Better | The agent does not quote prices, discounts or cost estimates for any plan. It may still describe the features and technical limits of each plan. When a customer asks about price, point them to the Sales team. | ### 2: Write by intent, not by keyword Guardrails match the **meaning** of a question, not a string of characters. A description along the lines of "block the word X" falls over the moment someone asks indirectly or asks in English. | | Description text | | - | - | | ❌ Weak | Block anything containing the word "competitor". | | ✅ Better | The agent does not compare our products with other vendors or comment on their strengths and weaknesses, even when asked indirectly. | Then pour a variety of phrasings into the **Samples** box, one per line: * "Is vendor X any better?" * "How does this compare with what we use today?" * "If it were you, who would you pick?" * "Why should we pick you over vendor X?" ### 3: Match the response type to the risk | Risk level | Response type to use | | - | - | | Low - off-topic subjects, small talk | **Agent writes its own reply**, so the wording fits the context | | High - legal, financial, medical, reputational crisis | **Fixed passage** signed off by Legal or Communications | A fixed passage only gets 300 characters, so cover two things: the reason for declining, and the next step. > I am not able to advise on legal matters. Please send your question to the Legal department inbox for direct support. Once written, paste the lines from your Samples box into the **Test** tab to check the rule catches them, then ask one ordinary question close to the subject to be sure it is not over-blocking. ## Things to keep in mind * **Description matters more than topic**: the model recognises situations from the Description, so spell out the scope instead of leaving a vague phrase. * **Use real examples**: take questions users have actually asked as Samples and the rule catches far more accurately. * **One topic per rule**: bundling several topics into one rule makes it hard to pause or edit any part of it later. * **Test before publishing**: use the **Test** or **Try it** tab to ask exactly the questions in Samples and watch how the agent reacts. * **Guardrails do not replace Instructions**: they block what must not happen; how to do the job right still belongs in the Instructions. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.