> ## Documentation Index
> Fetch the complete documentation index at: https://slackhive.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Test Mode

> Run sandboxed test sessions with MCP access and multi-agent delegation - before going live in Slack.

Test Mode lets you have a conversation with an agent inside the dashboard, exactly the way Slack would, but without posting anything to a real channel. MCPs work, boss delegation works, memory works - it's a safe sandbox for iterating.

## Open the Test tab

1. Go to the agent's detail page
2. Click **Test** in the tab bar
3. You'll see an empty conversation pane

## Start a test session

Type your prompt at the bottom and press Enter:

```
What tables do we have in the warehouse?
```

The agent streams its reply with the same tool-call trace you'd see in logs - every MCP call, every read, every tool result is visible inline.

## Use MCPs in the test

MCPs assigned to the agent work in Test Mode. If the agent has `playwright`, asking it to browse a URL actually opens a browser. If it has `notion`, it reads real Notion pages. No mocking - real tools, real side effects.

<Note>
  Side effects are real. Test deletes, writes, or postings are not isolated. Use a sandbox workspace for destructive MCPs.
</Note>

## Delegate to other agents

Type `@another-agent-slug` in your message and Test Mode routes the request through your Boss agent, exactly like Slack would. You'll see:

1. Your original message
2. Boss's delegation line
3. The specialist's reply streaming in
4. Boss's summary

This exercises the full multi-agent delegation path without sending anything to Slack.

## Inspect the trace

Each turn shows:

* **Tool calls** - chips with the tool name, args, and result
* **Cost** - total tokens and dollar estimate
* **Duration** - wall-clock time for the turn
* **Session ID** - click to tie back to the log stream

## When to use Test Mode

* **Before first deployment** - confirm the agent replies the way you expect
* **After a Coach-driven change** - sanity-check the new prompt
* **When an MCP seems off** - reproduce without the Slack round-trip
* **Debugging delegation** - exercise `@mention` chains in isolation

Once it works in Test Mode, it works in Slack.

## From tests to evals

When a test prompt represents behavior you never want to regress, turn it into an eval case:

1. Copy the exact prompt
2. Open agent → **Settings** → **Evals**
3. Add a regression case with checks for the expected answer or tool usage
4. Approve it and run the regression suite

See [Agent Evals](/agents/evals) for healthchecks, regression cases, verdicts, and judge settings.