Testing and Improving your Agent
Building a great AI agent is an iterative process. Cortex's test mode gives you a safe environment to run your agent, review its behaviour, and refine its instructions before anything goes live.
Test mode
Every new agent starts in test mode. While in test mode:
Read access is real. The agent reads live data from Street, Spectre, and Inspector, so it's working with real information and you can see how it handles actual records.
Write access is simulated. Any action the agent tries to take - sending an email, creating a record, updating a note - is simulated rather than carried out. You can see exactly what the agent would have done without any risk.
This means you can test your agent against real scenarios with zero risk of it accidentally contacting a client, altering a record, or taking any live action.
Watching your agent work
When you run an agent, Cortex shows you a real-time view of what it's doing. You can see:
Tool usage - Which tools the agent is calling and in what order.
Reasoning - The agent's thought process as it decides what to do next.
Outputs - The content it generates, such as draft emails, notes, or blog posts.
This transparency makes it easy to understand why the agent is behaving the way it is, which is essential for writing better instructions.
The feedback loop
After a test run completes, Cortex's built-in AI assistant will ask you how the agent performed. You have two options:
"It looks good"
If you're happy with the agent's behaviour, you can confirm and move on. When you're confident the agent is ready, you can switch it from test mode to live mode, where it will begin taking real actions.
"I want to improve it"
If something isn't right, you can describe what you want to change in plain English. For example:
"The tone of the emails is too formal - make it warmer and more conversational."
"It should not mention the asking price in the initial email."
"When there's no email address, it should create a note for the negotiator instead of skipping the record."
The AI assistant will suggest updates to your agent's instructions and tools. You can accept these suggestions as-is, edit them, or write your own changes. Then run the agent again and see how it performs.
The iterative process
The best agents are rarely perfect on the first run. A typical workflow looks like this:
Write your initial instructions and select tools.
Run the agent in test mode.
Review the outputs and reasoning.
Describe what you'd like to change.
Accept or edit the suggested improvements.
Run again.
Repeat until you're satisfied.
This cycle is fast and low-risk. Each iteration refines the agent's behaviour, and because test mode uses real data with simulated actions, you're always testing against realistic scenarios.
Going live
Once you're confident your agent is performing as expected, switch it to live mode. In live mode, the agent will carry out real actions — sending real emails, creating real records, and interacting with real data. Any tools with approvals configured will still require human sign-off before executing.
Street is constantly being updated. This can mean that some of these articles become slightly out of date. If you spot something that looks a little bit different to your Street account, please let us know on the live chat so we can get it updated!