RebelCore™ Agent

The RebelCore™ Agent is a prompt-driven companion to RebelCore™. Instead of clicking through forms, you describe what you want in natural language and the agent runs the underlying workflow for you.

It’s accessible at https://agent.rebelcore.ai and from the Open in RebelCore™ Agent button on specs in the dataset tree.

Governance — RAG with audit

The Agent is the inference tier of the Medallion architecture and acts as a governed RAG pipe over the Gold vectors. Analysts can prompt the Agent without ever seeing the raw or silver tiers underneath, and every prompt is audited (who asked, what gold context was used, what the Agent returned). See Governance & access for the full pattern.

When to use the agent

Common use cases:

• Build a spec — give the agent a spec and a target shape; it figures out the steps.
• Ask about your data — “what columns are common across these three sheets?”, “find rows where amount > 1000”.
• Run a multi-step task — anything that would normally take several manual steps, in one prompt.

If you can describe it in a sentence or two, try the agent first.

How a prompt session works

1. Open the agent (either directly or via the Open in RebelCore™ Agent button on a spec).
2. Choose Fast or Thinking before the first prompt in the session.
3. Type your prompt in the input box.
4. The agent runs the request — you’ll see status updates as it works through the steps.
5. The result appears in the response area. You can refine and re-prompt without losing context.

Tip

The more specific your prompt, the better the result. “Build a customer summary spec from sheet customers_q4, group by region, total by amount” is much more useful than “summarise this”.

Fast and Thinking

The toggle beside the dataset chips controls how a new session works:

Mode	Best for	What you see
Fast	Everyday questions, quick checks, and direct answers.	The answer streams back as it is written. Progress appears in the User Feedback tab.
Thinking	More complex, multi-step work where deeper traceability matters.	The agent runs the full governed workflow. Trace detail appears in the Reasoning tab.

The choice is locked after the first prompt in a session. To switch modes, start a new session. This keeps the conversation and audit trail consistent.

Fast still runs through governance: human-review checks, tool selection, dataset access, exposure limits, response masking, and audit all still apply. It is simply optimised for a shorter, streamed path.

Side panels

Beyond the prompt input, the Agent UI has dedicated panels that let you inspect and steer what’s going on:

• Sessions — your conversation history.
• Audit panel — the audit trail of what’s been asked and done in this project.

Picking datasets — the settings cog

The cog button below the chat input opens a popover where you tick the datasets (semantic specs) you want the agent to query. Each row shows the dataset name and a coloured pill — Full, Limited, or Advisory only — indicating the project’s LLM exposure level for that dataset.

When two or more datasets are ticked, a summary line under the list shows the effective exposure for this chat:

Effective for this chat: Advisory only (most restrictive of 2 levels)

The effective level is the most-restrictive across all picked datasets. The agent operates at this level for every prompt in the session. The selection persists across reloads of the agent.

The same pills appear on the chips at the very bottom of the chat box (next to the model badge), so the active level is visible regardless of whether the popover is open.

Dataset access — what you can see

The agent only lists datasets you have access to. A dataset (semantic spec) belongs to a project, and a project either:

• has All Users access — every user in the customer can use it, OR
• has Restricted Access — only listed users (plus the project owner) can use it.

If a project’s access setting changes (or you’re removed from its restricted list) and you had its dataset selected, the next time you open the agent the dataset is silently removed from your selection — your local selection is reconciled against what the server actually returns. This is enforced both at the list endpoint (the popover only shows datasets you can see) and at the inference endpoint (any unauthorised id you send is dropped before the workflow runs).

You won’t see datasets from projects in other customers, period.

Per-response indicator

Every assistant message has a small footer line — assistant · <timestamp> — that also carries a coloured pill showing the active LLM exposure level for that turn. If a project’s effective exposure changes mid-session (e.g. you toggled an advisory dataset on), the next response’s badge reflects the new level.

This badge isn’t decorative — it tells you what the agent was allowed to say. An Advisory-tagged response that reads like business advice is operating exactly as designed; a Full-tagged response that quotes specific values is also operating as designed.

Privacy - Mask to AI and Mask

The settings cog has a separate Privacy section with two controls:

• Mask to AI controls whether personal identifiers are replaced before the model sees the prompt, recent chat context, and tool results. When on, names, emails, phone numbers and regional ID numbers are sent as placeholders such as [PERSON_1] or [EMAIL_ADDRESS_1]. Business context such as amounts, dates, percentages and scores is kept so the Agent can still reason over the request.
• Mask controls what you see in the chat. When on, personal information remains masked in the response and history. When off, the server may reveal known placeholders back to you if your role and licence allow it.

Both controls default to ON. Mask to AI is fixed after the first prompt in a session because the session’s history and placeholder map must stay consistent. To change Mask to AI, start a new session.

Customer Super Admins can disable raw personal information being sent to AI for the whole customer. When that policy is off, Mask to AI is forced on and the Agent UI disables the checkbox.

What the agent can’t do

• The agent only acts on data already imported into your project. If you need new data, import it first.
• Some destructive actions (like deleting a project) require you to do them yourself in RebelCore™ — the agent will refuse.
• If a workflow takes too long or hits an error, you’ll see a clear message rather than a partial result.

Going back to RebelCore™

The agent has a Go to RebelCore™ button (top-left or in its menu) that takes you back to the main workspace. If you opened the agent from a specific spec, the link preserves the project context so you land back where you were.