Skip to main content
The Agent Sandbox is where you iterate on an AI agent against your own products, see exactly what it generates, and refine its behavior through a chat, all before you run enrichment on your catalog. It’s the fast, low-stakes workspace for tuning an agent’s voice, structure, and accuracy, so the agent is already performing the way you want before it touches any real job.

​
How it works

The sandbox lives inside the agent wizard’s final step, both when you create a new agent and when you edit an existing one. You pick a handful of real products from the current category, click Run, and the agent generates values for each one in front of you. For every result you can see the generated value, the AI’s reasoning, and the source data it used. The sandbox only runs when you click Run (or Re-run) — opening the step or switching attribute types doesn’t trigger a run on its own. That way you stay in control of when the LLM is called. While a run is in progress, the Run button becomes Cancel: click it to abort the in-flight request immediately, keep any results that already came back, and return to a startable state. If the output isn’t quite right, you don’t rewrite the prompt by hand. You tell the sandbox’s agent chat what to change (for example, “make descriptions shorter” or “don’t mention competitor brands”) and the chat rewrites the prompt as a new version in history. Each refinement creates a new version of the prompt that you can step through with undo and redo, so you can compare wordings and roll back any change that made things worse. Once the prompt is rewritten, click Re-run to generate fresh output for the new version. When the sandbox is showing output you’re happy with, save the agent. The current prompt is committed, and the agent is ready to be linked to attributes and run through enrichment at scale.

​
Why you would use this

Catch problems before they hit your catalog

See the agent’s output on real products first. Fix tone, structure, and accuracy on a handful of samples instead of discovering issues across thousands of enriched products.

Iterate in chat, not in the prompt editor

Tell the chat what to change in plain language. It rewrites the prompt and re-runs the agent for you, faster than hand-editing a long prompt and guessing at phrasing.

Compare versions with undo and redo

Step back through prompt versions to compare outputs. If a change made things worse, one click takes you back.

Use real products, not synthetic tests

The sandbox runs against your own product data, so what you see is what enrichment will actually produce.
Key concept: Agent chat. The chat panel on the left of the sandbox doesn’t just comment on the output. It rewrites the agent’s prompt based on what you ask and saves the new version in history. Click Re-run when you’re ready to generate output for the rewritten prompt — the chat itself doesn’t auto-trigger a new sandbox run.
The agent sandbox showing agent chat on the left with an assistant welcome message, and generated results for several products on the right with their images and outputs

​
Step-by-step guide

1

Open the sandbox

Create a new agent or edit an existing one. In the wizard, step through to the final step: Review & Test. The sandbox opens with a small random sample of products pre-selected for you.
For a brand new agent, the sandbox pre-selects up to 5 random products (preferring products with images), but it doesn’t run until you click Run. You can replace the selection any time before running.
2

Choose products to test

Use the product picker to set which products the agent should run against. You can:
  • Search and pick manually: type part of a product name, click results to add them
  • Click 5 Random: let Cernel pick 5 random products from the current category, preferring ones with images
You can select up to 10 products per run.
Pick a mix that represents what enrichment will actually face: different product types, price points, and levels of input data. Catching edge cases here saves you from discovering issues across thousands of enriched products later.
The sandbox product selector showing a searchable dropdown with product thumbnails, and a '5 Random' quick-pick button to the right
3

Run the agent

Click Run. The sandbox works through each product: analyzing configuration, processing products, running agent logic, and generating results. Each result shows the product’s name and image alongside the generated value for the agent’s attribute.On the next run, the button becomes Re-run, so you can generate fresh output on the same selection as you iterate. While a run is in progress, the same button shows Cancel — click it to stop the run immediately. Results that already came back are kept; you’ll see a Run cancelled toast and the button returns to Run / Re-run so you can adjust and try again.
Cancel is especially useful for External Data agents — a single agentic web research pass can take a couple of minutes, and Cancel lets you abort if you spot a misconfiguration without waiting it out.
4

Review the results

Scan the generated values. For any result you can:
  • Hover the value to see the AI Reasoning, a breakdown of how the agent arrived at the output
  • Click View details (the info icon at the top-right of each product card) to open the detail drawer, which shows the compiled Prompt the model received, the Source Data (product attributes used as context), and the Result in full
If a product failed, the card shows an error state so you know something went wrong for that specific input.
5

Refine through chat

If the output needs work, use the agent chat on the left. Type what you want to change. For example:
  • “Make the descriptions shorter, aim for 80 to 100 words.”
  • “Start every description with the material, then move to fit.”
  • “Don’t mention competitor brands.”
The chat rewrites the prompt and saves the change as a new version in history. To see the rewritten prompt in action, click Re-run — the chat doesn’t trigger a new sandbox run on its own, so you stay in control of when the LLM is called.
Type @ in the chat to reference a specific product from your current results. The chat picks up that product’s generated output and reasoning, so feedback like “@Blue Running Shoe, too wordy, make it match the tone of the others” is tied to exactly what you’re pointing at.
6

Step through versions with undo and redo

Every refinement creates a new version of the prompt. Use the Undo and Redo buttons at the top of the chat panel to move between versions. Each step takes you one change back or forward, so you can compare output side by side and roll back any change that didn’t land.
If the current direction isn’t working, undo back to a version you liked before adding new instructions on top. It’s often faster than undoing one refinement at a time.
7

Save the agent

When the output is where you want it, save the agent. The current prompt (the latest version in history) is committed and linked to the agent. Your sandbox selection and chat history are cleared; the agent is now ready to be linked to attributes and run through enrichment.
The agent is saved with the prompt you tested. Running enrichment with this agent will produce output similar to what you just saw in the sandbox.
The sandbox doesn’t run an agent’s dependencies before testing. If your agent depends on other attribute values (for example, a Meta Description agent that reads Product Description), pick products that already have those upstream values filled in. Otherwise the sandbox will show weaker output than enrichment would produce once dependencies are in place.

​
Advanced configuration

The agent chat supports @ mentions. Typing @ opens a list of the products in your current sandbox run. Click one to insert it as a pill into your message. When the chat rewrites the prompt, it uses that product’s generated output, reasoning, and images as grounded feedback, not just the text you typed.Use this when one product is off but the rest look fine. Instead of rewording the entire prompt, point the chat at the outlier and describe what’s wrong with it specifically.
The detail drawer ( View details on any result) shows three tabs:
  • Prompt: the compiled prompt that was sent to the model for that product, with Data Source references and product fields already filled in. If the output is surprising, check the compiled prompt first to see whether the right context made it in.
  • Source Data: the product attributes the agent used as context, including the product name, images, and any referenced fields
  • Result: the generated value on its own, rendered the way the product table will render it
If you’re debugging an unexpected output, comparing the Prompt and Source Data between a good product and a bad product usually pinpoints the problem quickly.
If you change the agent’s attribute type during the wizard (for example, from Text to HTML, or from Single Select to Multi Select), the sandbox automatically asks the chat to adapt the prompt for the new type. This keeps the instructions aligned with what the new type expects, so you don’t have to rewrite them from scratch.After the prompt is adapted, re-run the sandbox to see output in the new format.
  • Saved with the agent: the prompt as it stands at the moment you save (the current version in history), plus any configuration changes made in earlier wizard steps
  • Not saved: the list of sandbox products you tested on, individual sandbox results, and the agent chat transcript
When you reopen the wizard later, the sandbox starts fresh. The agent carries forward the prompt you saved, but the conversation that produced it is not stored.

​
Frequently asked questions

Two common causes:
  1. The agent has dependencies that haven’t been filled in on the test products. The sandbox doesn’t run upstream agents first. If your agent reads other attribute values, pick test products that already have those values.
  2. The selected products don’t represent your catalog well. Try clicking 5 Random again, or hand-pick a mix of product types and price points. Outputs can vary a lot between, say, a simple t-shirt and a technical running shoe.
No. The sandbox runs on up to 10 products per run by design, so refinement stays fast. Ten is typically enough to spot patterns in the output. Once you’re happy with the agent, run enrichment to generate content across the rest of the catalog.
Not directly. Undo and redo move one step at a time through the prompt’s history, so you can compare consecutive versions. If the current direction isn’t working, undo back to the version you liked and start a new refinement from there.
The chat calls the LLM once to rewrite your prompt for the new instruction — that’s a small, fast call. The sandbox itself only re-runs when you click Re-run, so you decide when the (potentially slow) enrichment pass actually fires. Keep chat messages focused: one clear change per message is easier to verify than a long instruction that rewrites several things at once.
Yes. The Run button becomes Cancel while a run is in progress. Click it to abort immediately — Cernel stops the in-flight LLM call, keeps any results that already came back, and returns the button to Run / Re-run so you can adjust and try again. This is especially useful for External Data agents in Agentic mode, which can take a couple of minutes per pass.
The agent chat transcript isn’t stored. The prompt version you saved with the agent is kept, but the conversation that produced it is discarded. If you reopen the agent, the sandbox starts fresh.
Technically yes. You can send chat messages before the sandbox has run, and the chat will still refine the prompt. But the refinements aren’t grounded in any output, so the results are less predictable. The recommended flow is: run the sandbox, read the output, then refine.

​
What’s next

AI Agents

Learn the full agent setup: prompts, Data Sources, dependencies, and linking agents to categories.

Your First Enrichment

Take a sandbox-tested agent and run it across a full category.

Attributes

Understand the attribute types the sandbox tests against, and how they’re configured on categories.

Data Sources

Plug brand guidelines and context libraries into prompts so sandbox output lines up with your brand.