CREATE AGENT Statement in Confluent Cloud

Confluent Intelligence enables creating Streaming Agents using declarative SQL syntax. Agents can reason over streaming data, invoke tools, and produce structured outputs.

Syntax

Standard agent

CREATE AGENT [IF NOT EXISTS] [catalog.][database.]agent_name
  USING MODEL <model_identifier>
  USING PROMPT <prompt_string>
  [USING TOOLS <tool1>, <tool2>]
  [COMMENT <comment_string>]
  WITH (<option_list>)

Reflection agent

CREATE AGENT [IF NOT EXISTS] [catalog.][database.]agent_name
  USING AGENTS <agent1>, <agent2>
  [COMMENT <comment_string>]
  WITH (<option_list>)

Description

Create a new Streaming Agent that can process streaming data using LLM reasoning and tool invocation capabilities.

A standard agent uses a specified model, prompt, and optional tools to process input data.

A reflection agent is a composite agent that orchestrates two sub-agents (a drafter and a critic) in an iterative loop to produce higher-quality output. The reflection agent does not have its own model, prompt, or tools. For more information, refer to Improve Agent Output with Reflection Workflows.

Parameters

model_identifier (STRING): Reference to a registered model (standard agents only)
prompt_string (STRING): System prompt for the agent (standard agents only)
tool1, tool2 (STRING): Comma-separated list of tool names (standard agents only)
agent1, agent2 (STRING): References to registered sub-agents (reflection agents only). The first agent is the drafter, and the second is the critic.
comment_string (STRING): Optional comment describing the agent

Agent options

Standard agent options

max_consecutive_failures: Maximum consecutive failures (optional, default: 3)
max_iterations: Maximum iterations for calling tools (optional, default: 10)

Reflection agent options

TYPE: Agent type, must be 'REFLECTION' (required)
PASS_CONDITION: Termination condition for the reflection loop, 'APPROVAL' or 'CONFIDENCE' (optional, default: APPROVAL)
MAX_ITERATIONS: Maximum reflection loop iterations (optional, default: 10)
CONFIDENCE_VALUE: Confidence threshold, 0.0-1.0, used when PASS_CONDITION is 'CONFIDENCE' (optional, default: 0.8)

Examples

Basic agent

CREATE AGENT weather_agent
  USING MODEL openai
  USING PROMPT 'Find weather info for provided city'
  USING TOOLS mcp_server, get_weather_tool
  WITH (
    'max_iterations' = '5'
  );

Complete example

-- Create MCP server connection
CREATE CONNECTION mcp_connection
WITH (
  'type' = 'mcp_server',
  'api-key' = '<your-api-key>',
  'endpoint' = 'https://mcp.example.com',

);

-- Create tools
CREATE TOOL mcp_server
USING CONNECTION mcp_connection
WITH (
  'type' = 'mcp',
  'allowed_tools' = 'tool1,tool2',
  'request_timeout' = '30'
);

-- Register a function.
CREATE FUNCTION convert_to_celsius
USING JAR 'celsius.jar'
COMMENT 'function to convert degree to celsius';

-- Create a tool based on the function.
CREATE TOOL convert_to_celsius_tool
USING FUNCTION convert_to_celsius
WITH (
  'type' = 'function',
  'description' = 'This function can be used by model to convert degree to celsius'
);

-- Create a connection to the model provider.
CREATE CONNECTION openai_connection
WITH (
  'type' = 'openai',
  'endpoint' = 'https://api.openai.com/v1/chat/completions',
  'api-key' = 'your-openai-key'
);

-- Register a remote model.
CREATE MODEL openai
INPUT(text STRING)
OUTPUT (res STRING)
WITH (
  'provider' = 'openai',
  'openai.model' = 'gpt-4o',
  'openai.connection' = 'openai_connection'
);

-- Create an agent that uses the model and tools.
CREATE AGENT weather_agent
USING MODEL openai
USING PROMPT 'Find weather info for provided city'
USING TOOLS mcp_server, convert_to_celsius_tool;

Reflection agent example

-- Create the drafter agent.
CREATE AGENT claim_extractor
USING MODEL claims_model
USING PROMPT 'You are a claims intake agent. Extract a JSON
  object with fields: incident_type, damage_summary,
  estimated_severity (low/med/high), estimated_cost, location
  (lat/long), and any missing_fields.';

-- Create the critic agent.
CREATE AGENT claim_critic
USING MODEL claims_model
USING PROMPT 'You are a claims quality reviewer. Critique the
  extracted JSON for missing required fields, contradictions,
  or low-confidence assumptions.';

-- Create the reflection agent.
CREATE AGENT reflective_claim_intake
USING AGENTS claim_extractor, claim_critic
WITH (
  'type' = 'reflection',
  'pass_condition' = 'approval',
  'max_iterations' = '3'
);

Constraints

The following constraints apply to reflection agents:

You must not specify USING MODEL, USING PROMPT, or USING TOOLS on the reflection agent.
You must specify exactly two sub-agents in the USING AGENTS clause. The first agent is the drafter, and the second is the critic.