Skip to content

Tools in chat: confirmations and errors

To answer many questions, Intra AI uses tools: it searches the web, runs code, loads a skill or, through an agent, works in an integration such as Outlook or Google Drive. This page shows how to check what a tool did, how to approve or cancel a write action, and what to do when a tool call fails.

See what was called

Every time Intra AI uses tools, a tool card appears in the answer. Its header sums up what happened in one sentence, such as "Read 1 item" or "Read 3 emails in Outlook". While a call is still running, the header says what is happening right now, for example "Reading … in …".

On the right of the header you see the logos of the systems involved and the total duration. A green check mark means everything worked. A red icon and a note such as "1 call failed" mean at least one call did not succeed. An amber warning icon with "1 call did not complete" means a call was cut off. While a confirmation is waiting for you, the header reads "Waiting for your answer".

  1. Click the header. The card expands and lists each call with its duration.
  2. Click a call to see its result right inside the card.
A tool card above the answer shows what Intra AI called to produce it. Expand it to see each individual call.Enlarge figure

Check a call in the Tool inspector

The Tool inspector shows what went into each call and what came back. Use it when an answer doesn't match what you expected.

  1. In the tool card, click the Details icon on the far right, either in the header or next to a single call. Next to a call, the icon appears when you hover over the call. The inspector opens beside the conversation, or full screen on a phone.
  2. Use Previous step and Next step to move between calls. The top of the panel shows where you are, for example Step 1 of 3.
  3. Check Status (OK or Failed), Duration and Time. For many calls, Context also shows how much of the result the model actually read.
  4. Switch between the Result, Input and Raw tabs. Input shows the values Intra AI passed to the tool. Raw shows the unprocessed data.
  5. Use Copy or Download JSON to take the data with you, for example for a support request.
  6. Close the inspector with Close inspector (the X), or click the same icon again.
The Tool inspector shows the status, duration and time of each call, along with the input it received.Enlarge figure

When an agent brings in another agent to help, that agent gets its own card. Its Details button opens the inspector on the Trace tab, which lists that agent's rounds and calls along with Duration, Rounds, Calls and Tokens.

Approve or cancel a write action

Before an agent creates, changes, sends or deletes something, it stops and shows a confirmation card such as "Send this message?" or "Confirm this calendar change?". The card is marked Waiting for your answer. Actions that only read data never ask.

  1. Check the fields on the card, such as To, Subject and Body for an email, or Start, End and Attendees for a meeting. If the text is cut off, click Show full body.
  2. If everything is right, click the button named after the action, such as Send, Create entry, Apply change or Confirm action. Some cards simply show Approve. For a deletion, the Delete button is red.
  3. If something is wrong, click Cancel. The card then reads "Cancelled — nothing was written." Tell the agent what to do differently.

Cancel is selected by default, so pressing Enter by accident cancels the action and nothing is written. The action runs with your own connected account, never with the account of whoever built the agent. A card stays valid for 24 hours by default. After that it expires and shows "This request has expired." Ask the agent for the action again.

Ask first or run without asking

Whoever builds the agent decides whether a write action asks first. In the Tools step, under What the agent may do on its own, each write action is set to one of two modes:

Mode, What happens in chat
ModeWhat happens in chat
always ask before runningThe confirmation card appears, and nothing happens until you approve.
run without askingThe action runs right away. Afterward the call in the tool card is marked Ran without confirmation, and the card opens by itself so you can see what was written.

New actions start as please decide. On a shared agent, you have to set each one before you can finish the agent. If you pick run without asking, write down in the field below the action why it may skip the confirmation. The agent must also already have a procedure, and the builder currently can't create one. Without a procedure the agent can't be finished, so set the action to always ask before running instead.

Warning: The mode run without asking applies to everyone who uses a shared agent. They work with their own accounts and don't see a card either. Keep new actions on "always ask" until a test run shows the agent fills in the action correctly.

Your instance can be set up centrally so that agents never write without asking, or so that specific tools always ask. The confirmation card then appears even for an action set to run without asking. To give actions to an agent, see Tools, knowledge and sub-agents.

When tools are available

In a conversation without an agent, Intra AI only uses its built-in tools: web search when you turn it on, document search, your skills and, if your instance offers it, the code interpreter. Tools from integrations like email, calendars or file storage are used by an agent that has been given those actions.

Many of these integrations connect through MCP (Model Context Protocol), an open standard that lets an AI application call tools in other systems.

A tool is available in chat when the agent has been given the action and all of these are true:

  • Your administrator has set up the integration on your instance.
  • Your administrator has approved that specific tool. If its definition changes, Intra AI holds it back until it is approved again.
  • Your administrator has not limited the integration to groups or tenants you don't belong to.
  • You have connected your own account under Integrations. If you haven't, a notice above the message box says the agent "needs access to" the integration and offers Connect. If it says "Your administrator needs to set this up" instead, there is nothing you can do yourself.

When a tool was held back, the answer shows the notice "Some tools were unavailable." The answer was written without that tool and may be incomplete. Ask your administrator to review the tools that were held back.

Troubleshooting

Expand the tool card: the failed call states the reason there. Click the call, and below it there is often an ID such as "Reference: …". Include it when you ask your administrator for help, because it identifies that exact call.

Message starts with, What to do
Message starts withWhat to do
"The … connection expired"Open Integrations, sign in to the service again and repeat your question.
"The call to … failed and was not repeated"Check in the target system whether the action went through anyway, for example in your Sent folder. Only then run it again.
"… returned a result too large to use"Narrow the request, for example to a date range, a folder or a sender.
"The tool did not respond in time", "The connected system is rate limiting requests", "The tool service is busy"Wait a moment and try again.
"… could not be reached", "… is temporarily unavailable"Try again later. If it keeps happening, contact your administrator.
"The … integration is not configured", "… is not configured on this instance"Only your administrator can fix this.
"This write is not delegated to the specialist"The agent that was brought in isn't allowed to run this action itself. Usually the calling agent then does the write. To change this for good, your instance must allow sub-agents to write (off by default), and whoever maintains the agent must grant write permissions in the advanced settings under Write rights for the sub-agent.
"The model sent malformed arguments to …"Nothing. Intra AI retries on its own.

Note: Never repeat a write action with an unclear outcome right away. You could send an email twice or create a duplicate meeting.