Agent Runtime Preview
AI funding and usage
AI included uses REDA credits. Bring your own AI uses your organisation’s provider account, billed directly by that provider. The licence selects one funding model for the organisation. In either model, REDA AI governs configured tools and actions in Salesforce. Set up your licence and AI accounts.
With AI included, select the agent’s Processing Tier. With Bring your own AI, select AI Connection and Model in the agent’s AI Account section, or inherit the organisation default. Different agents can use different supported providers and models. Prompts, tools, record context and channel permissions still define each agent’s job. Configure the agent’s AI account.
What the Preview Shows
The preview reads the agent definition, applies the same channel and context-object filters the runtime applies, and assembles four panels:
| Panel | What it lists |
|---|---|
| Prompts | Active custom prompts — agent-specific and shared — in the order they will be assembled. Each prompt shows its type, sort order, content preview and a token estimate, plus a flag if it doesn’t fit within the agent’s token budget. |
| Tools | The read-only tools the agent will see (built-in plus custom). Filtered by channel scope and context-object relevance, with each tool’s tier and category. |
| Actions | The Action tools the agent will see, with confirmation requirements visible at a glance. |
| Suggestions | The suggestion chips that will appear on the empty state of the chat, filtered to match the channel and context object. |
A header bar above the panels shows the agent identity, the resolved channel (the one being previewed), the context object (if any), and the configured token budget.
Configuration Conflicts & Health Score
Alongside the four panels, the preview runs a deterministic conflict-detection pass over the agent’s configuration. The result is two things:
- A health score from 0 to 100, summarising how clean the configuration is.
- A list of individual conflicts classified as High (likely to break the agent) or Warning (worth reviewing).
Conflict detection is free — it doesn’t call the model. The intent is to catch the obvious mistakes — missing prompts, unassigned tools that look orphaned, channel mismatches, contradictory instructions — before they show up in real conversations.
Channel & Context Object Switcher
The preview accepts two parameters that change what gets loaded:
Channel
Defaults to the agent’s configured channel, but you can switch it to any other channel the agent might run on. The preview re-filters tools, prompts and suggestions to match. Use this to confirm that:
- Tools you intended for the API channel are actually scoped that way
- Prompts you wrote for an external widget aren’t accidentally leaking into internal conversations
- Suggestion chips you authored for the community channel show up there and not on the External Widget
Context Object
Defaults to no record context. Set the API name of an object to preview what the agent loads when running on a record page of that object. The preview applies the same context-object relevance scoring the runtime uses, so tools that are scoped to specific objects move in and out of the visible list as you change the value.
For agents deployed on multiple objects, switching the context object reveals how the agent looks “from” each one — useful when one record page surfaces fewer tools than another.
Token Budget Tracking
The Prompts panel keeps a running total of estimated tokens as you scroll. Each prompt is flagged green or red against the agent’s configured token budget; prompts that push the running total above the budget are visibly out of bounds.
This is an estimate — the runtime tokeniser is the source of truth — but it’s usually accurate enough to catch the case where a recently-added prompt has blown the budget. Pair it with the Conversation Debugger after a real run to confirm.
Read-Only and Cheap
- No data changes when you browse. Opening the preview only reads data (no DML, Salesforce's term for record inserts/updates/deletes). The one deliberate exception is Edit Loadout: from the preview an admin can exclude specific tools, prompts or suggestion chips from this agent's loadout, and saving writes those exclusions to the Agent Definition record in one atomic update — you need edit permission on the agent to use it.
- No model call. The preview doesn’t consume credits. The conflict detector is purely deterministic Apex.
- Runs as the current user. The preview queries data (via SOQL, Salesforce's query language) using the running user’s sharing rules and field-level security (FLS). If a prompt or tool isn’t visible to the user previewing it, it won’t appear — reflecting what that user would actually get at runtime.
Where to Find It
Open the preview from any Agent Definition record — the runtime preview component is embedded in the agent’s admin layout. You can also open it directly from the REDA AI Admin tab.
How It Differs from the Conversation Debugger
The Runtime Preview shows what the agent will load on its next conversation; the Conversation Debugger shows what the agent did load on a specific past conversation, including the model’s actual tool calls and responses.
- Use the Runtime Preview while building or refining an agent — before any conversation happens.
- Use the Conversation Debugger when something went wrong in a real conversation and you want to see why.
Per-Agent Loadout (Edit Loadout)
Some agents need a curated subset of the tools, actions, prompts and suggestions that the package and your customizations make available. A common case: a Community-channel portal agent that uses custom “elevated” tools (running in without sharing mode) to reach data that the portal user cannot access directly. In that case the standard user-mode tools (e.g. create_record, update_record, soql_query) overlap with the elevated equivalents, and you want the LLM to see only one of them. Similar needs arise for prompts (e.g. exclude a verbose persona prompt for a token-budget-tight agent) and for suggestion chips (hide a suggestion that does not apply to this agent).
The Edit Loadout action on the Runtime Preview applies to all four categories — Tools, Actions, Prompts, Suggestions — and lets administrators exclude specific items from a given agent without touching shipped Custom Metadata, custom prompt records, or affecting any other agent in the org. ISV-shipped (universal) and client-defined items are equally excludable.
Exclusions are stored on the Agent Definition record as three separate lists — one each for excluded tools/actions, excluded prompts, and excluded suggestions. Tools and Actions share the same exclusion list because they are managed together under the hood. At runtime, REDA AI filters each list before assembling the agent’s system prompt, tool catalogue and suggestion chip set — excluded items never reach the LLM, but stay visible (greyed out, with an Excluded badge) in the Runtime Preview so admins can see and re-include them.
Using Edit Loadout
- Open the Agent Definition record where the Runtime Preview component is placed.
- Switch to any tab — Prompts, Tools, Actions or Suggestions. The Edit Loadout button appears in the toolbar at the top of every tab.
- Click Edit Loadout. A checkbox appears next to every item across all four tabs.
- Uncheck any item you want to exclude for this agent. Excluded items are greyed out and tagged with an Excluded pill. The summary in the toolbar updates as you go.
- Move freely between tabs — your unsaved changes persist across tabs while in edit mode.
- Click Save Loadout to persist all four lists in a single transaction, or Cancel to discard your changes.
Exclusions apply only to the specific agent you are editing. Other agents in the org — including a different community agent or any internal agent — continue to see all items they would normally load. Newly-installed package items always default to active on every agent: you opt them out, not in. A package upgrade that ships a brand-new tool, prompt or suggestion will not silently disappear for any agent.
The exclusion lists are meant to be maintained through the Edit Loadout action. Editing them directly on the record works but bypasses the validation and normalization (sorting, deduplication, blank trimming) that Edit Loadout applies. Use Edit Loadout instead.
When to exclude an item
- You replaced a standard tool with a custom equivalent — e.g. a
create_record_elevatedrunningwithout sharingfor portal users. Exclude the standardcreate_recordon that agent so the LLM does not pick it. - The agent should never perform a class of action — e.g. a read-only research agent that should never be able to
delete_recordorupdate_record. - A universal prompt does not fit a specific agent — e.g. a verbose “safety guardrails” prompt that is universal but redundant for an internal-only agent that already has tighter rules in its persona. Exclude it to recover its token budget.
- A suggestion chip does not apply — e.g. a generic “Show my open cases” suggestion that is irrelevant to a sales-focused agent. Exclude it so portal users do not see a chip the agent cannot answer.
- You want to minimise the agent’s context size — fewer items = smaller LLM context = lower credit consumption per turn. For agents with a narrow purpose, dropping unused items is a quick win.
- An item is producing unwanted behaviour — exclude it temporarily while you investigate, then re-include.
Best Practices
Open it after every agent edit. The preview is the cheapest way to verify that a change had the effect you expected — especially when you’re wiring channel or context-object scoping for the first time.
Walk through every channel the agent supports. An agent that lists API and Internal as channels should be previewed under both. Tools and prompts often look right on one channel and wrong on another.
Use it during demos. Showing the preview to a stakeholder is a fast way to communicate what an agent does without setting up a live conversation.
Triage conflicts before testing. The configuration conflict list catches issues that would otherwise burn a real conversation. Clear the High items and review the Warnings before you run a single message through the agent.
Treat token estimates as a yardstick. An estimate is not a guarantee — for production cost decisions, run a real conversation and read the credits actually consumed.
Configure reusable business instructions
Exclusive Prompts let administrators assign reusable instructions to selected agents. Review the Exclusive Prompt flag and Summary on Custom Prompt, then the assignment's Agent Definition, Custom Prompt, Active and Load Mode. Use the intended load mode; On Demand needs the context-prompt tool in the agent's loadout. Record-type brackets use shared scope rules, including in preview and prompt administration. Prompt and tool ordering is deterministic, and runtime preview can use an actual record ID. Updated language instructions follow the message on API/external channels; user placeholders apply only where the authenticated user context is intended. Preserve custom prompts when merging the current defaults.
Frequently Asked Questions
No. It’s pure Apex — no model calls. Open it as often as you like.
Check three things: the channel you’re previewing must be in the tool’s Channels list; the agent name on the tool must match (or be blank); and Is Active must be on. Switching channels in the preview is usually the fastest way to find which scope is filtering it out.
The preview estimates are character-based heuristics. The runtime tokeniser is the source of truth. The estimates are accurate enough to catch obvious budget overruns but won’t match the runtime exactly.
Yes. The preview runs under the running user’s sharing rules and field-level security, so a prompt or tool that one user can’t see won’t appear in their preview. This mirrors what the runtime would actually deliver to that user.
The health score reflects the conflict detector’s findings. Lower scores mean more or higher-severity conflicts. Read the conflict list for specifics — the score itself is a summary, not the diagnosis.
Yes. Exclusions are stored on the individual Agent Definition record, so they only affect the agent you edited. Other agents — including a different community agent or any internal agent — continue to see the full catalogue of tools, actions, prompts and suggestions. There is no global blocklist.
Intentional. The preview lists every item (tool, action, prompt, suggestion) that could load for the agent so you can see what is excluded and re-include it later. Excluded items are greyed out and tagged with an Excluded pill. At runtime, the LLM never sees them — REDA AI strips them before the system prompt and tool catalogue are sent.
The Runtime Preview focuses on prompts, tools, actions, suggestions and conflicts. For context-field details, see Understanding Context Fields and the Context Field Optimizer; for the actual context the agent received in a past conversation, use the Conversation Debugger.
Next Steps
Read Build Agent with AI for AI-assisted improvements to an agent’s configuration, or revisit Configuring Agent Settings for the underlying agent fields the preview reflects.