Tool Loading & Cost
How Tool Loading Works
On every model call, the platform assembles the agent’s toolbox from the active Tool Definitions and sends it with the request. A tool makes the list when it passes all of these filters:
- Active — the Tool Definition is marked active.
- Agent scope — built-in tools are available to every agent; a custom tool can be scoped to a specific agent via its Agent field.
- Channel — the tool’s Channels selection includes the channel the conversation runs on.
- Context — if the tool names Context Objects, the current record’s object must match.
- Not excluded — the tool is not on this agent’s exclusion list (set via Edit Loadout in the Agent Runtime Preview).
- Within Max Tools — the agent’s Max Tools setting (1–100, default 25) caps how many tools are offered, in Sort Order.
The resulting list is the same on every turn of the conversation — and, for a given agent and channel, the same across conversations. That stability is the point.
Why a Stable Toolbox Is Cheap
A stable list of permitted tools helps the provider reuse cached prompt content where its API and model support caching. Definitions still count toward the request’s tokens. Cache handling and charges vary by provider and model; use the usage record and provider console rather than assuming a fixed discount.
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.
The Real Cost Levers
| Lever | Where | What it does |
|---|---|---|
| Don’t over-assign | Max Tools on the Agent Definition; Edit Loadout exclusions | The cheapest tool is one the agent doesn’t carry. Trim the toolbox to what the agent actually uses. |
| Tools Mode | Agent Definition | Record Pages Only skips loading tools on home and list pages, where the agent has no record to act on. |
| Tight schemas | Custom Tool Definitions | A verbose input schema costs tokens on every turn and produces worse tool calls. Keep descriptions crisp and fields minimal. |
| Keep the list stable | Configuration discipline | Every change to an agent’s toolbox invalidates the provider cache for that agent. Batch tool changes rather than tweaking daily. |
See Credit Usage Optimisation for the broader set of cost techniques — token budgets, history budgets, processing tiers and caching.
The Legacy Tool Tier Field
Tool Definitions still carry a Tool Tier field. It is deprecated: retained so existing configurations keep saving cleanly, but ignored by the runtime in every current version. Setting it changes nothing. If you documented internal runbooks around Tier 1 / Tier 2 behaviour, retire them — the loading model above is the whole story.
Best Practices
- Audit large schemas. If a custom tool’s input schema is verbose, ask whether it needs to be. A tighter schema is cheaper on every turn and produces better tool calls.
- Don’t over-assign. Assigned tools cost tokens on every turn. The biggest saving is a smaller toolbox, via Max Tools and Edit Loadout exclusions.
- Verify with the Conversation Debugger. The Debugger shows the exact tool list the model saw on each turn — use it to confirm the toolbox is what you intended (it is the same list on every turn).
Frequently Asked Questions
A stable list of permitted tools helps the provider reuse cached prompt content where its API and model support caching. Definitions still count toward the request’s tokens. Cache handling and charges vary by provider and model; use the usage record and provider console rather than assuming a fixed discount.
No. Any Tier values you set are simply ignored. Your cost levers today are Max Tools, per-agent tool exclusions (Edit Loadout in the Agent Runtime Preview), Tools Mode, and keeping custom tool schemas tight.
Trim the toolbox: lower Max Tools or exclude tools the agent never uses (Edit Loadout), set Tools Mode to Record Pages Only if the agent only works on records, tighten verbose custom tool schemas, and keep the tool list stable — every change to the list invalidates the provider cache for that agent.
No. Tool loading is a token-economy mechanism, not a security boundary. Every tool runs with CRUD, field-level security and sharing applied (subject to the agent’s Execution Context on the API channel). To restrict a tool from a channel or an agent, use the Channels and Agent fields on the Tool Definition.
Next Steps
Read Credit Usage Optimisation for the full set of techniques to keep credit consumption low, or revisit the Built-in Tools Reference to see each tool’s inputs and channels.