Agent Abilities
An agent is only as useful as what it can do. In Achilles, each specific ability is called a tool. This page explains the kinds of abilities, how they are kept safe, and how to choose the right ones.
Four kinds of things an agent can do
| Ability | What it is | Example |
|---|---|---|
| Look up records | Reads information from your systems. | "Show my renewals this week." |
| Take an action | Changes something or sends something. Needs approval. | "Send this renewal reminder." |
| Give a quick answer | Handles a common request instantly. | "What is my policy status?" |
| Work through several steps | Combines lookups and decisions to finish a task. | Check claim status, explain the missing document, offer to notify the customer. |
How an ability is defined
Every ability has a plain description, the details it needs, and what it returns. For example, "Look up a policy's renewal date" needs a policy number and returns the date, premium and status.
You do not need to write these by hand. Agent Builder creates them from your SOP, and you can review them under Tools in the console. Good descriptions matter, because the agent uses them to decide what to do. See Agent Builder.
How abilities are checked
Every ability passes through safety checks that run in the platform, not in the AI's own reasoning.
- Is this person allowed to use it? Roles decide. See Roles and access.
- Are the details valid? Missing or wrong details are caught before anything runs.
- Which records may they see? Results are limited to what that person is allowed to see.
- Does it need approval? Actions that change things wait for a yes. See Approvals.
- Is it recorded? Every use is logged. See Audit trail.
The AI never does the arithmetic. Premiums, balances and dates come from your systems, and the agent only words the answer. That keeps numbers accurate.
Quick answers for common questions
Most real traffic is a handful of repeated questions. For these, Achilles can recognise the request and go straight to the right lookup. This is called the fast path. Customers get answers sooner and your costs are lower.
Rare or complicated requests still use the full multi-step approach. You do not have to choose. The platform picks the quickest safe route.
Open-ended questions
Sometimes people ask things you did not plan for, such as "How many policies did my branch renew in March?" You can allow the agent to answer these from your data with a dynamic query ability.
It stays controlled:
- It can only read, never change.
- It can only use the tables you approve.
- Sensitive fields are kept out of reach.
- Every question is limited to the records that person may see.
Prefer a named ability whenever one fits. Use the open-ended option for the long tail.
Choosing abilities
| Question to ask | Why |
|---|---|
| Does the agent really need it for this job? | Fewer abilities mean fewer mistakes. |
| Is it read-only? | Start with lookups and add actions later. |
| Who should be allowed to use it? | Give it to specific roles only. |
| What could go wrong? | Add approval or handoff for risky actions. |
Add actions one at a time. Test each one thoroughly in Playground before moving on. See Test your agent.