Tuning your agent

Agents get better when they're shaped around your real workflow. Here's what makes the biggest difference.

Instructions & tone

• Write clear instructions for what the agent should and shouldn't do
• Give a few examples in your voice so replies sound like you
• State the rules explicitly — refunds, escalation, anything sensitive

Knowledge

• Connect your real docs, FAQs, and data so answers are grounded
• Keep sources clean and current — quality in, quality out
• Require citations for anything fact-based

Behaviour controls

• Lower the temperature for focused, consistent replies; raise it for more varied, creative output
• Set when the agent escalates to a human
• Cap response length and add guardrails for off-topic or unsafe requests

Handling common requests

• Look at the questions that come up most and make sure each is handled cleanly
• Add the ones it gets wrong to its knowledge or examples
• Iterate from real conversations — small, frequent tweaks beat big rewrites