AgentServer

The AgentServer wraps a KAOS Agent in a FastAPI application with OpenAI-compatible chat API, A2A discovery, memory endpoints, and Kubernetes probes.

Class Definition

class AgentServer:
    def __init__(self, agent: Agent, settings: AgentServerSettings)

Endpoints

Path	Method	Description
/health	GET	Kubernetes liveness probe
/ready	GET	Kubernetes readiness probe
/.well-known/agent.json	GET	A2A agent card discovery
/v1/chat/completions	POST	OpenAI-compatible chat (streaming + non-streaming)
/memory/events	GET	List memory events
/memory/sessions	GET	List memory sessions

POST /v1/chat/completions

OpenAI-compatible chat endpoint. Supports both streaming and non-streaming:

# Non-streaming
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "agent", "messages": [{"role": "user", "content": "Hello!"}]}'

# Streaming
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "agent", "messages": [{"role": "user", "content": "Hello!"}], "stream": true}'

GET /.well-known/agent.json

A2A-compliant agent card (v0.3.0) with name, description, skills (discovered from MCP tools), capabilities, and version:

curl http://localhost:8000/.well-known/agent.json

GET /memory/events

curl http://localhost:8000/memory/events?session_id=abc&limit=50

AgentServerSettings

All configuration via environment variables:

class AgentServerSettings(BaseSettings):
    agent_name: str                    # AGENT_NAME (required)
    model_api_url: str                 # MODEL_API_URL (required)
    model_name: str                    # MODEL_NAME (required)
    agent_description: str = "AI Agent"
    agent_instructions: str = "You are a helpful assistant."
    agent_port: int = 8000
    agent_log_level: str = "INFO"
    agent_access_log: bool = False

    # Sub-agents
    agent_sub_agents: str = ""         # "name:url,name:url"
    peer_agents: str = ""              # "worker-1,worker-2" (K8s format)

    # MCP servers
    mcp_servers: str = ""              # "echo,calc"
    # + MCP_SERVER_ECHO_URL, MCP_SERVER_CALC_URL

    # Agentic loop
    agentic_loop_max_steps: int = 5

    # Memory
    memory_enabled: bool = True
    memory_type: str = "local"
    memory_context_limit: int = 6
    memory_max_sessions: int = 1000
    memory_max_session_events: int = 500
    memory_redis_url: str = ""

Sub-Agent Formats

Direct format (AGENT_SUB_AGENTS):

export AGENT_SUB_AGENTS="worker-1:http://worker-1:8000,worker-2:http://worker-2:8000"

Kubernetes format (PEER_AGENTS + individual URLs):

export PEER_AGENTS="worker-1,worker-2"
export PEER_AGENT_WORKER_1_CARD_URL="http://worker-1:8000"
export PEER_AGENT_WORKER_2_CARD_URL="http://worker-2:8000"

MCP Server Configuration

export MCP_SERVERS="echo,calc"
export MCP_SERVER_ECHO_URL="http://echo-mcp:8000"
export MCP_SERVER_CALC_URL="http://calc-mcp:8000"

URLs have /mcp auto-appended for Streamable HTTP transport.

Factory Functions

get_app

Lazy factory for uvicorn with --factory:

uvicorn pais.server:get_app --factory --host 0.0.0.0 --port 8000

create_agent_server

from pais.server import create_agent_server, AgentServerSettings

# From environment variables
server = create_agent_server()

# With explicit settings
settings = AgentServerSettings(
    agent_name="my-agent",
    model_api_url="http://ollama:11434",
    model_name="llama3.2",
)
server = create_agent_server(settings)

Lifecycle

The server manages agent lifecycle via FastAPI lifespan:

@asynccontextmanager
async def _lifespan(self, app: FastAPI):
    yield
    await self.agent.close()

OpenTelemetry

When OTEL_EXPORTER_OTLP_ENDPOINT is set, the server automatically instruments:

• HTTP server spans (FastAPI)
• HTTP client spans (outgoing requests)
• Custom agent spans (tool calls, delegation)

export OTEL_SERVICE_NAME="my-agent"
export OTEL_EXPORTER_OTLP_ENDPOINT="http://jaeger:4318"