Agent Instructions

Agent instructions define how an AI agent should behave while pursuing a goal. They establish its role, priorities, limits, allowed methods, output format, and situations in which it must ask for approval or stop.

What instructions control

Instructions may specify:

• The agent’s responsibility and audience.
• Which sources or tools it may use.
• Required steps or quality checks.
• Safety, privacy, and permission boundaries.
• Preferred tone or output structure.
• When to retry, escalate, or request clarification.
• Which actions require human approval.

For example, a research agent might be instructed to use primary sources, separate verified facts from interpretation, cite important claims, and avoid publishing automatically.

Instructions differ from goals

A goal describes the desired result. Instructions describe the operating rules.

• Goal: Prepare a competitor comparison for a product decision.
• Instructions: Compare equivalent capabilities, use current official documentation, state uncertainty, and do not include unverified pricing.

Both are components of the system described in Anatomy of an AI Agent.

Instructions come from several layers

An agent may receive system-level instructions from its developer, task instructions from a user, procedural guidance from a reusable skill, and constraints from the surrounding application.

These layers can conflict. A well-designed system establishes a clear priority order and prevents lower-trust content, such as a retrieved webpage, from overriding trusted instructions. This is especially important for reducing prompt-injection risk.

Instructions are not guaranteed behavior

Instructions guide a language model; they do not mechanically enforce every rule. Ambiguous wording, long context, conflicting directions, or adversarial content can cause failures.

Critical constraints should therefore be backed by system controls. Permissions can block unauthorized tools. Schemas can reject invalid arguments. Approval gates can stop sensitive actions. Deterministic validators can check outputs before they are used.

Good instructions are operational

Useful instructions are specific enough to guide decisions but not filled with unnecessary prose. They define observable behavior, clear boundaries, and recovery paths.

The strongest agent design combines concise instructions with tools, state, permissions, and tests. Prompt wording alone should not carry the full burden of safety or reliability.