cidadao.ai-backend / src /services /agent_orchestrator.py
anderson-ufrj
feat(orchestration): implement advanced agent orchestration system
88b8ba0
"""
Advanced Agent Orchestrator for Cidadão.AI.
Manages complex agent coordination patterns and workflows.
"""
import asyncio
import time
from datetime import datetime, timedelta
from typing import Any, Dict, List, Optional, Callable, Tuple
from dataclasses import dataclass, field
from enum import Enum
from collections import defaultdict
import weakref
from src.core import get_logger
from src.agents.deodoro import BaseAgent, AgentContext, AgentMessage, AgentResponse
from src.services.agent_lazy_loader import agent_lazy_loader
from src.services.cache_service import CacheService
from src.core.exceptions import OrchestrationError
logger = get_logger("agent.orchestrator")
class OrchestrationPattern(Enum):
"""Orchestration patterns supported by the system."""
SEQUENTIAL = "sequential"
PARALLEL = "parallel"
FAN_OUT_FAN_IN = "fan_out_fan_in"
CONDITIONAL = "conditional"
SAGA = "saga"
MAP_REDUCE = "map_reduce"
EVENT_DRIVEN = "event_driven"
class CircuitState(Enum):
"""Circuit breaker states."""
CLOSED = "closed"
OPEN = "open"
HALF_OPEN = "half_open"
@dataclass
class CircuitBreaker:
"""Circuit breaker for agent fault tolerance."""
failure_threshold: int = 5
recovery_timeout: int = 60 # seconds
half_open_requests: int = 3
state: CircuitState = CircuitState.CLOSED
failure_count: int = 0
last_failure_time: Optional[datetime] = None
success_count: int = 0
@dataclass
class WorkflowStep:
"""Represents a step in an orchestrated workflow."""
step_id: str
agent_name: str
action: str
input_mapping: Dict[str, str] = field(default_factory=dict)
output_mapping: Dict[str, str] = field(default_factory=dict)
conditions: Dict[str, Any] = field(default_factory=dict)
retry_config: Dict[str, Any] = field(default_factory=dict)
timeout: int = 300 # seconds
@dataclass
class WorkflowDefinition:
"""Defines an orchestrated workflow."""
workflow_id: str
name: str
pattern: OrchestrationPattern
steps: List[WorkflowStep]
metadata: Dict[str, Any] = field(default_factory=dict)
timeout: int = 1800 # 30 minutes
@dataclass
class OrchestrationMetrics:
"""Metrics for orchestration performance."""
total_executions: int = 0
successful_executions: int = 0
failed_executions: int = 0
total_duration_seconds: float = 0.0
agent_execution_times: Dict[str, List[float]] = field(default_factory=lambda: defaultdict(list))
pattern_usage: Dict[str, int] = field(default_factory=lambda: defaultdict(int))
class EventBus:
"""Simple event bus for event-driven choreography."""
def __init__(self):
self._handlers: Dict[str, List[Callable]] = defaultdict(list)
self._async_handlers: Dict[str, List[Callable]] = defaultdict(list)
def on(self, event_name: str, handler: Callable):
"""Register an event handler."""
if asyncio.iscoroutinefunction(handler):
self._async_handlers[event_name].append(handler)
else:
self._handlers[event_name].append(handler)
async def emit(self, event_name: str, data: Any = None):
"""Emit an event to all registered handlers."""
event = {"name": event_name, "data": data, "timestamp": datetime.utcnow()}
# Call sync handlers
for handler in self._handlers.get(event_name, []):
try:
handler(event)
except Exception as e:
logger.error(f"Error in event handler: {e}")
# Call async handlers
tasks = []
for handler in self._async_handlers.get(event_name, []):
tasks.append(handler(event))
if tasks:
await asyncio.gather(*tasks, return_exceptions=True)
class AgentOrchestrator:
"""Advanced orchestrator for multi-agent coordination."""
def __init__(self):
self.logger = logger
self._workflows: Dict[str, WorkflowDefinition] = {}
self._circuit_breakers: Dict[str, CircuitBreaker] = {}
self._metrics = OrchestrationMetrics()
self._event_bus = EventBus()
self._cache = CacheService()
self._agent_capabilities: Dict[str, List[str]] = {}
self._running_workflows: weakref.WeakValueDictionary = weakref.WeakValueDictionary()
async def initialize(self):
"""Initialize the orchestrator."""
self.logger.info("Initializing Agent Orchestrator")
# Discover agent capabilities
await self._discover_agent_capabilities()
# Register default workflows
self._register_default_workflows()
async def _discover_agent_capabilities(self):
"""Discover capabilities of all available agents."""
try:
agents = await agent_lazy_loader.list_agents()
for agent_info in agents:
agent = await agent_lazy_loader.get_agent(agent_info["name"])
if hasattr(agent, 'capabilities'):
self._agent_capabilities[agent_info["name"]] = agent.capabilities
except Exception as e:
self.logger.error(f"Error discovering agent capabilities: {e}")
def _register_default_workflows(self):
"""Register default workflow patterns."""
# Investigation workflow
investigation_workflow = WorkflowDefinition(
workflow_id="default_investigation",
name="Standard Investigation Workflow",
pattern=OrchestrationPattern.SEQUENTIAL,
steps=[
WorkflowStep(
step_id="anomaly_detection",
agent_name="zumbi",
action="detect_anomalies"
),
WorkflowStep(
step_id="pattern_analysis",
agent_name="anita",
action="analyze_patterns",
conditions={"if": "anomalies_found", "gt": 0}
),
WorkflowStep(
step_id="report_generation",
agent_name="tiradentes",
action="generate_report"
)
]
)
self._workflows["default_investigation"] = investigation_workflow
async def execute_workflow(
self,
workflow_id: str,
initial_data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute a complete workflow."""
workflow = self._workflows.get(workflow_id)
if not workflow:
raise OrchestrationError(f"Workflow {workflow_id} not found")
self.logger.info(f"Executing workflow: {workflow.name}")
start_time = time.time()
try:
# Track metrics
self._metrics.total_executions += 1
self._metrics.pattern_usage[workflow.pattern.value] += 1
# Execute based on pattern
if workflow.pattern == OrchestrationPattern.SEQUENTIAL:
result = await self._execute_sequential(workflow, initial_data, context)
elif workflow.pattern == OrchestrationPattern.PARALLEL:
result = await self._execute_parallel(workflow, initial_data, context)
elif workflow.pattern == OrchestrationPattern.FAN_OUT_FAN_IN:
result = await self._execute_fan_out_fan_in(workflow, initial_data, context)
elif workflow.pattern == OrchestrationPattern.CONDITIONAL:
result = await self._execute_conditional(workflow, initial_data, context)
elif workflow.pattern == OrchestrationPattern.SAGA:
result = await self._execute_saga(workflow, initial_data, context)
elif workflow.pattern == OrchestrationPattern.MAP_REDUCE:
result = await self._execute_map_reduce(workflow, initial_data, context)
else:
raise OrchestrationError(f"Unsupported pattern: {workflow.pattern}")
# Update metrics
duration = time.time() - start_time
self._metrics.successful_executions += 1
self._metrics.total_duration_seconds += duration
return {
"workflow_id": workflow_id,
"status": "completed",
"result": result,
"duration": duration
}
except Exception as e:
self._metrics.failed_executions += 1
raise OrchestrationError(f"Workflow execution failed: {e}")
async def _execute_sequential(
self,
workflow: WorkflowDefinition,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute workflow steps sequentially."""
current_data = data.copy()
results = []
for step in workflow.steps:
# Check conditions
if not self._check_conditions(step.conditions, current_data):
continue
# Execute step
step_result = await self._execute_step(step, current_data, context)
results.append(step_result)
# Map output to next input
for output_key, data_key in step.output_mapping.items():
if output_key in step_result.get("data", {}):
current_data[data_key] = step_result["data"][output_key]
return {
"pattern": "sequential",
"steps_executed": len(results),
"final_data": current_data,
"step_results": results
}
async def _execute_parallel(
self,
workflow: WorkflowDefinition,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute workflow steps in parallel."""
tasks = []
for step in workflow.steps:
if self._check_conditions(step.conditions, data):
tasks.append(self._execute_step(step, data.copy(), context))
results = await asyncio.gather(*tasks, return_exceptions=True)
return {
"pattern": "parallel",
"steps_executed": len(results),
"results": [r for r in results if not isinstance(r, Exception)],
"errors": [str(r) for r in results if isinstance(r, Exception)]
}
async def _execute_step(
self,
step: WorkflowStep,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute a single workflow step."""
start_time = time.time()
try:
# Get agent
agent = await self._get_agent_with_circuit_breaker(step.agent_name)
# Prepare input data
input_data = {}
for input_key, data_key in step.input_mapping.items():
if data_key in data:
input_data[input_key] = data[data_key]
# Create message
message = AgentMessage(
type=step.action,
data=input_data or data,
sender="orchestrator",
metadata={"workflow_step": step.step_id}
)
# Execute with timeout
response = await asyncio.wait_for(
agent.process(message, context),
timeout=step.timeout
)
# Track metrics
duration = time.time() - start_time
self._metrics.agent_execution_times[step.agent_name].append(duration)
return {
"step_id": step.step_id,
"agent": step.agent_name,
"success": response.success,
"data": response.data,
"duration": duration
}
except asyncio.TimeoutError:
raise OrchestrationError(f"Step {step.step_id} timed out after {step.timeout}s")
except Exception as e:
raise OrchestrationError(f"Step {step.step_id} failed: {e}")
def _check_conditions(self, conditions: Dict[str, Any], data: Dict[str, Any]) -> bool:
"""Check if conditions are met for step execution."""
if not conditions:
return True
# Simple condition evaluation
if "if" in conditions:
field = conditions["if"]
if field not in data:
return False
value = data[field]
if "eq" in conditions:
return value == conditions["eq"]
elif "gt" in conditions:
return value > conditions["gt"]
elif "lt" in conditions:
return value < conditions["lt"]
elif "in" in conditions:
return value in conditions["in"]
return True
async def _get_agent_with_circuit_breaker(self, agent_name: str) -> BaseAgent:
"""Get agent with circuit breaker protection."""
circuit_breaker = self._circuit_breakers.get(agent_name)
if not circuit_breaker:
circuit_breaker = CircuitBreaker()
self._circuit_breakers[agent_name] = circuit_breaker
# Check circuit state
if circuit_breaker.state == CircuitState.OPEN:
# Check if recovery timeout has passed
if (datetime.utcnow() - circuit_breaker.last_failure_time).seconds > circuit_breaker.recovery_timeout:
circuit_breaker.state = CircuitState.HALF_OPEN
circuit_breaker.success_count = 0
else:
raise OrchestrationError(f"Circuit breaker open for {agent_name}")
try:
agent = await agent_lazy_loader.get_agent(agent_name)
# Reset on success
if circuit_breaker.state == CircuitState.HALF_OPEN:
circuit_breaker.success_count += 1
if circuit_breaker.success_count >= circuit_breaker.half_open_requests:
circuit_breaker.state = CircuitState.CLOSED
circuit_breaker.failure_count = 0
return agent
except Exception as e:
# Update failure count
circuit_breaker.failure_count += 1
circuit_breaker.last_failure_time = datetime.utcnow()
if circuit_breaker.failure_count >= circuit_breaker.failure_threshold:
circuit_breaker.state = CircuitState.OPEN
raise
async def select_best_agent(self, required_capabilities: List[str]) -> Optional[BaseAgent]:
"""Select the best agent based on required capabilities."""
best_match = None
best_score = 0
for agent_name, capabilities in self._agent_capabilities.items():
# Calculate capability match score
score = sum(1 for cap in required_capabilities if cap in capabilities)
if score > best_score:
best_score = score
best_match = agent_name
if best_match:
return await agent_lazy_loader.get_agent(best_match)
return None
async def execute_with_retry(
self,
agent: BaseAgent,
message: AgentMessage,
context: AgentContext,
max_retries: int = 3,
backoff_multiplier: float = 2.0,
fallback_agent: Optional[BaseAgent] = None
) -> AgentResponse:
"""Execute agent with retry logic and optional fallback."""
last_error = None
for attempt in range(max_retries + 1):
try:
return await agent.process(message, context)
except Exception as e:
last_error = e
if attempt < max_retries:
wait_time = (backoff_multiplier ** attempt) * 1.0
await asyncio.sleep(wait_time)
continue
# Try fallback agent if available
if fallback_agent:
self.logger.warning(f"Primary agent failed, trying fallback")
return await fallback_agent.process(message, context)
raise
raise OrchestrationError(f"All retry attempts failed: {last_error}")
def configure_retry_policy(self, policy: Dict[str, Any]):
"""Configure global retry policy."""
self._retry_policy = policy
def configure_circuit_breaker(self, config: Dict[str, Any]):
"""Configure circuit breaker settings."""
self._circuit_breaker_config = config
async def execute_conditional_workflow(
self,
workflow_def: Dict[str, Any],
initial_data: Dict[str, Any],
context: AgentContext
) -> List[Dict[str, Any]]:
"""Execute a conditional workflow with branching."""
execution_path = []
current_step = workflow_def["start"]
current_data = initial_data.copy()
while current_step:
step_def = workflow_def["steps"][current_step]
# Execute step
agent_name = step_def["agent"]
agent = await agent_lazy_loader.get_agent(agent_name)
message = AgentMessage(
type="process",
data=current_data,
sender="conditional_workflow",
metadata={"step": current_step}
)
response = await agent.process(message, context)
execution_path.append({
"step": current_step,
"agent": agent_name,
"success": response.success,
"data": response.data
})
# Determine next step based on conditions
next_step_def = step_def.get("next")
if not next_step_def:
break
if isinstance(next_step_def, str):
current_step = next_step_def
else:
# Conditional branching
condition = next_step_def.get("condition")
if condition == "anomalies_found":
if response.data.get("anomalies_detected", 0) > 0:
current_step = next_step_def.get("true")
else:
current_step = next_step_def.get("false")
elif condition == "high_risk":
if response.data.get("risk_level") == "high":
current_step = next_step_def.get("true")
else:
current_step = next_step_def.get("false")
else:
current_step = next_step_def.get("default")
# Update data for next step
current_data.update(response.data)
return execution_path
async def discover_agents(self) -> List[Dict[str, Any]]:
"""Discover all available agents."""
return await agent_lazy_loader.list_agents()
async def find_agents_with_capability(self, capability: str) -> List[Dict[str, Any]]:
"""Find agents with a specific capability."""
matching_agents = []
for agent_name, capabilities in self._agent_capabilities.items():
if capability in capabilities:
agent_info = {
"name": agent_name,
"capabilities": capabilities
}
matching_agents.append(agent_info)
return matching_agents
async def execute_with_circuit_breaker(
self,
agent: BaseAgent,
message: AgentMessage,
context: AgentContext
) -> AgentResponse:
"""Execute agent with circuit breaker protection."""
agent_name = agent.name
circuit_breaker = self._circuit_breakers.get(agent_name)
if not circuit_breaker:
circuit_breaker = CircuitBreaker(**self._circuit_breaker_config)
self._circuit_breakers[agent_name] = circuit_breaker
if circuit_breaker.state == CircuitState.OPEN:
if (datetime.utcnow() - circuit_breaker.last_failure_time).seconds < circuit_breaker.recovery_timeout:
raise OrchestrationError(f"Circuit breaker open for {agent_name}")
else:
circuit_breaker.state = CircuitState.HALF_OPEN
try:
response = await agent.process(message, context)
if circuit_breaker.state == CircuitState.HALF_OPEN:
circuit_breaker.success_count += 1
if circuit_breaker.success_count >= circuit_breaker.half_open_requests:
circuit_breaker.state = CircuitState.CLOSED
circuit_breaker.failure_count = 0
return response
except Exception as e:
circuit_breaker.failure_count += 1
circuit_breaker.last_failure_time = datetime.utcnow()
if circuit_breaker.failure_count >= circuit_breaker.failure_threshold:
circuit_breaker.state = CircuitState.OPEN
raise
async def start_saga(
self,
saga_definition: Dict[str, Any],
initial_data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Start a saga transaction."""
saga_state = {
"saga_id": f"saga_{datetime.utcnow().timestamp()}",
"name": saga_definition["name"],
"current_step": 0,
"completed_steps": [],
"compensated_steps": [],
"data": initial_data,
"completed": False,
"failed": False
}
return saga_state
async def execute_saga_step(
self,
saga_state: Dict[str, Any],
step: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute a saga step with compensation support."""
try:
agent_name = step["agent"]
agent = await agent_lazy_loader.get_agent(agent_name)
message = AgentMessage(
type=step["service"],
data=saga_state["data"],
sender="saga_coordinator",
metadata={"saga_id": saga_state["saga_id"]}
)
response = await agent.process(message, context)
if response.success:
saga_state["completed_steps"].append({
"step": step,
"result": response.data
})
saga_state["current_step"] += 1
if saga_state["current_step"] >= len(saga_state.get("total_steps", [])):
saga_state["completed"] = True
else:
saga_state["failed"] = True
# Trigger compensation
await self._compensate_saga(saga_state, context)
except Exception as e:
saga_state["failed"] = True
saga_state["error"] = str(e)
await self._compensate_saga(saga_state, context)
return saga_state
async def _compensate_saga(self, saga_state: Dict[str, Any], context: AgentContext):
"""Compensate completed saga steps."""
for completed_step in reversed(saga_state["completed_steps"]):
try:
step = completed_step["step"]
if "compensation" in step:
agent = await agent_lazy_loader.get_agent(step["agent"])
compensation_message = AgentMessage(
type=step["compensation"],
data={
"original_data": saga_state["data"],
"step_result": completed_step["result"]
},
sender="saga_compensator",
metadata={"saga_id": saga_state["saga_id"]}
)
await agent.process(compensation_message, context)
saga_state["compensated_steps"].append(step)
except Exception as e:
self.logger.error(f"Compensation failed for step: {e}")
def get_event_bus(self) -> EventBus:
"""Get the event bus for choreography."""
return self._event_bus
async def get_stats(self) -> Dict[str, Any]:
"""Get orchestrator statistics."""
return {
"total_executions": self._metrics.total_executions,
"successful_executions": self._metrics.successful_executions,
"failed_executions": self._metrics.failed_executions,
"success_rate": (
self._metrics.successful_executions / self._metrics.total_executions
if self._metrics.total_executions > 0 else 0
),
"average_duration": (
self._metrics.total_duration_seconds / self._metrics.successful_executions
if self._metrics.successful_executions > 0 else 0
),
"pattern_usage": dict(self._metrics.pattern_usage),
"agent_performance": {
agent: {
"executions": len(times),
"avg_time": sum(times) / len(times) if times else 0,
"min_time": min(times) if times else 0,
"max_time": max(times) if times else 0
}
for agent, times in self._metrics.agent_execution_times.items()
},
"circuit_breakers": {
agent: {
"state": cb.state.value,
"failure_count": cb.failure_count
}
for agent, cb in self._circuit_breakers.items()
}
}
async def _execute_fan_out_fan_in(
self,
workflow: WorkflowDefinition,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute fan-out/fan-in pattern."""
# Fan-out: execute multiple steps in parallel
fan_out_results = await self._execute_parallel(workflow, data, context)
# Fan-in: aggregate results
aggregated_data = {
"pattern": "fan_out_fan_in",
"fan_out_results": fan_out_results["results"],
"aggregated_data": {}
}
# Simple aggregation - can be customized
for result in fan_out_results["results"]:
if result.get("success") and "data" in result:
aggregated_data["aggregated_data"].update(result["data"])
return aggregated_data
async def _execute_map_reduce(
self,
workflow: WorkflowDefinition,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute map-reduce pattern."""
# Map phase
map_items = data.get("items", [])
map_results = []
for item in map_items:
# Execute map step for each item
map_step = workflow.steps[0] # Assuming first step is map
step_result = await self._execute_step(
map_step,
{"item": item},
context
)
map_results.append(step_result)
# Reduce phase
reduce_step = workflow.steps[1] # Assuming second step is reduce
reduce_data = {
"map_results": [r["data"] for r in map_results if r.get("success")]
}
reduce_result = await self._execute_step(
reduce_step,
reduce_data,
context
)
return {
"pattern": "map_reduce",
"map_count": len(map_results),
"reduce_result": reduce_result["data"] if reduce_result.get("success") else None
}
async def _execute_saga(
self,
workflow: WorkflowDefinition,
data: Dict[str, Any],
context: AgentContext
) -> Dict[str, Any]:
"""Execute saga pattern with compensation."""
saga_state = await self.start_saga(
{"name": workflow.name, "steps": workflow.steps},
data,
context
)
saga_state["total_steps"] = workflow.steps
for step in workflow.steps:
saga_state = await self.execute_saga_step(
saga_state,
{"agent": step.agent_name, "service": step.action},
context
)
if saga_state["failed"]:
break
return {
"pattern": "saga",
"saga_id": saga_state["saga_id"],
"completed": saga_state["completed"],
"failed": saga_state["failed"],
"completed_steps": len(saga_state["completed_steps"]),
"compensated_steps": len(saga_state["compensated_steps"])
}
# Global orchestrator instance
orchestrator = AgentOrchestrator()
async def get_orchestrator() -> AgentOrchestrator:
"""Get the global orchestrator instance."""
return orchestrator