Customizing watsonx Orchestrate agents with plug-ins unlocks powerful ways to shape how agents interpret inputs and produce outputs. By integrating plug-ins that filter, sanitize, or transform data, teams can enforce critical safety, guardrails, security, and regulatory requirements. This post explores how input and output plug-ins strengthen agent reliability and trustworthiness.

Niklas Heidloff

Watsonx Orchestrate now supports plug-ins that intercept agent inputs and outputs to enforce safety, security, and compliance requirements. Two plug-in types exist: pre-invoke (runs before the agent processes a request) and post-invoke (runs after the agent generates a response). Practical examples show how to mask or block offensive words in user input using AGENTPREINVOKE, and how to redact email addresses in agent output using AGENTPOSTINVOKE. Both plug-in types are implemented as Python tools with specific kind annotations, registered in the agent's YAML definition, and imported via the orchestrate CLI. Plug-ins also have access to user context such as email, tenant ID, and username.

Watsonx Orchestrate Plug‑ins for Security and Compliance