Best practices

Things that separate a working agent from a good agent. Each principle has its own page; read the ones relevant to what you’re building.

System instructions

• Be specific about the role and responsibilities.
• Include examples of desired behavior — show, don’t just tell.
• Set clear boundaries: what the agent should refuse, what it should escalate.
• Define communication tone and style explicitly.

See Writing good instructions.

Tools

• Enable only what the agent needs. More tools = higher cost, slower responses, and more chances for misuse.
• Test tools individually before chaining them.
• TeamMate injects only the built-in tools that fit each agent’s model and enabled capabilities, so you rarely need to micro-manage the tool list.

See Choosing: model vs skill vs tool.

Knowledge base

• Connect only relevant sources for the agent’s purpose.
• Keep KBs updated — stale content produces confidently wrong answers.
• Remove outdated or irrelevant sources.

See Knowledge base hygiene.

Testing

• Cover various input types before publishing.
• Verify tool usage works as expected.
• Check answers match your spec.
• Test edge cases and error scenarios.

See Debugging behavior.

Cost

• Pick the smallest model that meets the quality bar.
• Cap max tokens to discourage rambling.
• Use skills over inline tools when you’ll reuse them across agents.

See Cost-aware agent design.

Anti-patterns

See Common anti-patterns for the most frequent mistakes — overstuffed instructions, vague descriptions, missing guardrails, untested edge cases.