Instruction
Instructions serve as reusable blocks of system prompt logic. They guide the agent’s reasoning in each step by setting expectations, tone, or rules of engagement for the LLM. Defined independently, they are referenced by name and can be updated without touching agent configurations. An agent references a particular version of the instruction. Each instruction includes the following:
- A unique ID (ins_abcd)
- A name and description. The name must be unique across all instructions per customer
- A prompt template
- Optional metadata for versioning or rollout control
An agent references a particular version of the instruction. The version number is optional but can be specified in the request, or the latest version is selected. You must update the agent refer to any newer versions of instructions.
Instructions are the system-level configuration for agent behavior. They use the Velocity templating engine to allow for dynamic content.
Shared Instructions
Shared instructions are referenced by ID across multiple agents. You can define instructions independently on inline. Each instruction includes:
- A unique ID following the pattern ins_[identifier]
- A name that must be unique across all instructions per customer
- A description of the instruction's purpose
- A prompt template with support for variables populated from tool context
- Version management for controlled updates
- Optional metadata for categorization
When an instruction is updated, the changes create a new version. Agents reference a specific version of the instruction, so updates are not automatically reflected until the agent is explicitly updated to reference the newer version. The version number is optional in requests but can be specified for precise control. If omitted, the latest version is used.
Inline Instructions
Inline instructions have a one-to-one relationship with the agent and cannot be used with other agents. They are automatically deleted when the agent is deleted. This model provides simplified management for agent-specific behavioral logic that doesn't need to be shared.
Template Context
Instructions use the Velocity templating engine with access to tool context during execution. The template receives information about available tools and their execution results. Agent and session context are not included in the template scope for this version.
Instructions are the system-level configuration for agent behavior and provide the flexibility to define consistent reasoning patterns across different agents or create specialized behavior for specific use cases.
1