Tools

Tools are the capability layer. They make specific operations in your other systems available to an agent, under policy and access control, so it can retrieve or verify something instead of relying on what someone said or what it inferred.

Tools are how an agent gets past conversation and analysis into doing something useful with your systems.

What they are used for

Retrieval. Fetching authoritative information to ground an answer: account details, availability, case status, pricing rules, entitlement, history.

Enrichment. Assembling the relevant context for a conversation, such as recent activity, prior decisions and open items, within approved scopes.

Validation. Checking a constraint, verifying an identifier, confirming the current state before proposing anything.

Bounded operations. Explicitly permitted operations, such as looking up allowed options or checking required fields before an update.

Tools are exposed through toolkits, which centralise authentication, credentials, access control and audit. That is what keeps integrations governable: the agent asks for a capability, and your rules decide whether it happens.

How they are constrained

Enabling tools for an agent should mean:

• Explicitly scoped. Which systems, and which operations within them.
• Bound to policy. What data is reachable, which fields are off limits, what needs approval, what gets logged.
• Observable. Every invocation traceable to an agent, a session and a conversation.
• Skill-mediated. Skills decide when a tool is appropriate and how its result is used.

Tools and actions

The two are complementary and usually sequential. Tools ground and validate. Actions execute.

1. The agent uses tools to fetch the context relevant to the record and the conversation.
2. It produces summaries and signals anchored to what was actually said.
3. Based on those, it proposes or initiates actions against your systems.
4. Everything is recorded with provenance, so an outcome can be traced back to the conversation that caused it.

What makes a toolkit enterprise-grade

Only approved operations exposed. Access that respects tenant, team and record boundaries. Credentials managed centrally, rotated, and granted at least privilege. And a clear record of what was accessed, when, by whom, and under which version of the agent.

What you see when one runs

In a chat, a meeting or Ask Auron, each tool the agent calls appears in the thread as its own row, on both web and mobile: the tool's name, a spinner while it runs, a tick when it returns. The row sits where the call happened, between the text written before it and the text written after, so a reply that pauses to look something up reads as two blocks with the lookup in between rather than one paragraph that appears to change subject. The rows are labels only; nothing opens when you tap one.

Where this lives in the product

Toolkits are registered and attached to agents in the console under Toolkits. Connection secrets belong in Secrets so they can be rotated in one place.