Core configuration
Every agent has three required fields: Role sets the agent’s persona and decision-making frame."You are a senior financial analyst specializing in SaaS metrics."
Goal defines what the agent is trying to accomplish at a high level.
"Help finance teams understand their ARR, churn, and LTV from raw data exports."
Instructions contain detailed behavioral rules: tone, constraints, output format, and step-by-step logic. Better instructions produce better agent behavior.
Optional features
| Feature | What it does |
|---|---|
| Memory | Retains context across turns (session) or sessions (long-term via Cognis) |
| Knowledge Base | Connects RAG, Knowledge Graph, or Semantic Model retrieval |
| Tools | Lets the agent call external systems (APIs, apps, databases) |
| Responsible AI | Applies guardrail policies (PII, toxicity, injection blocking) |
| Global Context | Applies org-wide instructions to all agents in the workspace |
| Structured Output | Forces a JSON response conforming to a schema you define |
Agent as API endpoint
Every agent in Lyzr is automatically exposed as a REST endpoint at agent creation time. No server code is required. Available endpoint modes:POST /v3/agent/{id}/chat: standard chat (request/response)POST /v3/agent/{id}/stream-chat: streaming response (server-sent events)POST /v3/agent/{id}/multimodal-chat: accepts image and file inputs
session_id for multi-turn conversations. A new UUID in each request creates a stateless, one-off interaction.