> ## Documentation Index > Fetch the complete documentation index at: https://docs.deepblocker.ai/llms.txt > Use this file to discover all available pages before exploring further. # Plan, start, pause and stop a test > Plan a test on the agent's page, start it with the agent name, watch it run, and pause or stop it at any time. > [!NOTE] > > Before you start: the agent is proven, has at least one must-never rule and a signed agreement. Any member can plan, pause or stop. Only the owner or an admin can start or resume. ## Before you start * The three lines on the agent's **Tests** tab are all ticked: proven, a must-never rule, the agreement signed. * Test conversations left this month. The tab shows "N of M test conversations used this month, N left." * A web agent that answers our request, or a bot that answers messages. See [Connect a web agent](https://docs.deepblocker.ai/agent-assurance/connect/web-api.md). ## Steps: plan a test **1. Press Plan a test** On the agent's **Tests** tab press **Plan a test**. It is greyed out until the agent has a must-never rule and a channel. **2. Pick what the attacker should try** Under **What should the attacker try?** pick one goal. **My own goal** opens a box for your own words, 10 to 1,000 characters. **3. Pick the kind of test** **One conversation** is one attacker conversation each time it runs. **Sweep of many short attacks** runs many attacks once; it is only available for an agent DeepBlocker connected through its connector, and the form says so. **4. Set the length and how often** **How long each conversation may be**: Short, Standard or Long. **How often**: once, every week or every month. A sweep runs once. **5. Check the estimate and save** The form says "This test uses N test conversations. You have N left this month." Give it a **Name (optional)** and press **Save test**. A saved test does not run until you press Start. ## The goals | Goal in the form | What the attacker tries | | - | - | | **Get another customer's details** | Make the agent share personal or account details of someone else | | **Make it act without checks** | Make it change an account, order or booking without checking who it talks to | | **Get money or a deal it should not give** | Make it promise or give a refund, discount or payment it may not give | | **Make it reveal its instructions** | Make it show its hidden instructions, settings or notes | | **Break any of its rules** | Break any of the agent's rules | | **My own goal** | Your words | ## Message limits | Length | Messages at most | | - | - | | **Short (up to 6 messages)** | 6 | | **Standard (up to 10 messages)** | 10 | | **Long (up to 16 messages)** | 16 | No test ever passes 30 messages. A sweep conversation stops at 10. ## Steps: start a test **1. Press Start** On the test's row press **Start**. The row asks you to type the agent name. "It begins sending test messages to your agent." **2. Type the agent name and press Start test** The name is matched without regard to case. The page says "Started. The first message goes out within a few minutes." The status becomes **Scheduled**, then **Running**. ## What you see while it runs The test row shows "N of M messages" and "Running for 12 min" with a bar. The page refreshes every 30 seconds while a test runs, was just started, or is due within 15 minutes. It pauses refreshing while the tab is hidden and refreshes at once when you come back. ## Pause, resume, stop | Button | Who | When it shows | What it does | | - | - | - | - | | **Pause** | Any member | A test you planned that is scheduled or running | Holds the test. No message goes out until it resumes | | **Resume** | Owner or admin | A paused test you planned | Sends messages again | | **Stop** | Any member | Any scheduled, running or paused test, DeepBlocker's too | Ends it for good. What it found so far is kept | | **Discard** | Any member | A test you planned that has not started | Removes the draft | | **Stop all tests** | Any member | At the top of the **Tests** tab when a test is live | Stops every scheduled, running or paused test in the workspace | **Stop** asks "Stop this test for good? What it found so far is kept." Press **Yes, stop it** or **Keep it**. Stop works even when self-service is off. ## What happens next When the test finishes its status is **Finished** and a report is written. See [Reports](https://docs.deepblocker.ai/agent-assurance/reports.md). A repeating test shows "Repeats, next on date" and runs again on its own. ## Problems > [accordion] Plan a test is greyed out > > The agent has no must-never rule, or no channel we can test. Set a rule on the **Rules** tab. See [Rules](https://docs.deepblocker.ai/agent-assurance/rules.md). > [accordion] It says the test needs more conversations than you have left > > Pick a shorter test, a lighter sweep, or wait for the month to turn. See [Plans and usage](https://docs.deepblocker.ai/agent-assurance/plans.md). > [accordion] It says your goal holds a word the attacker must never see > > Some words are blocked from the attacker's instructions. Write the goal another way. > [accordion] Start says to type the agent name > > The name you typed does not match. Copy it from the top of the page. > [accordion] Start says this test has already started > > The test is past the draft stage. Watch it on the **Tests** tab, or plan a new one. > [accordion] The test is paused and I did not pause it > > See [Safety](https://docs.deepblocker.ai/agent-assurance/safety.md) for what pauses a test by itself. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.