Originally published on tamiz.pro.
The era of "chatbox" AI is giving way to autonomous agents that execute multi-step workflows, call external APIs, and manage state across complex decision trees. As these systems move from prototype to production, the primary challenge shifts from model accuracy to system reliability. Unlike traditional microservices, where failure modes are predictable (timeouts, 503s), autonomous agents suffer from non-deterministic failures: hallucinated tool arguments, context window exhaustion, and semantic drift. This article explores how to adapt traditional distributed systems patterns—specifically the Circuit Breaker, Retry with Jitter, and Bulkhead patterns—to the unique constraints of Large Language Model (LLM) interactions.
In traditional software engineering, we assume that if a function receives valid inputs, it will either produce a valid output or raise a specific exception. In agentic AI, this assumption breaks down. An agent interacting with an LLM is a stochastic system wrapped in a deterministic orchestration layer.
The "Unreliability Paradox" arises because we demand deterministic business outcomes (e.g., "book a flight") from non-deterministic underlying components (the LLM's token generation). When an LLM hallucinates a tool argument, the downstream API fails. In a standard system, we would catch the 400 error and flag it as a bug. In an agentic system, we must catch it, analyze the failure, and potentially retry the LLM with different instructions or context.
Standard distributed system patterns are necessary but insufficient. A standard HTTP retry does not help if the LLM consistently generates invalid JSON for a specific complex prompt. We need a higher-level abstraction of resilience: Semantic Resilience.
To design a fault-tolerant autonomous agent, we must map traditional patterns to the LLM context:
A traditional circuit breaker opens when a service returns 5xx errors or times out. For LLMs, the "circuit" should trip on semantic degradation or cost explosion, not just HTTP errors. If the LLM starts generating nonsense repeatedly for a specific task type, the circuit should open to prevent wasting tokens and propagate a fallback.
LLM providers have rate limits (RPM/TPM). However, simple backoff is dangerous for agents. If an agent fails at step 4 of a 10-step workflow due to a transient network error, retrying step 4 is safe. If it fails due to a context window overflow, retrying without modifying the context will fail again. Therefore, retries must be context-aware.
Isolate resources. If you are running multiple agents, you must isolate their LLM connections. A burst of traffic from one agent type should not starve another. This is achieved via connection pooling limits specific to agent classes.
A semantic circuit breaker monitors the "quality" of LLM outputs, not just their availability. It tracks metrics like:
Below is a Python implementation of a SemanticCircuitBreaker that wraps an LLM client. It tracks consecutive failures and opens the circuit if the failure rate exceeds a threshold within a rolling window.
import time
from enum import Enum
from dataclasses import dataclass, field
from typing import Optional
import asyncio
class CircuitState(Enum):
CLOSED = "closed"
OPEN = "open"
HALF_OPEN = "half_open"
@dataclass
class CircuitBreakerConfig:
failure_threshold: int = 5
recovery_timeout: float = 60.0
success_threshold: int = 2
window_size: int = 10
@dataclass
class SemanticCircuitBreaker:
config: CircuitBreakerConfig
state: CircuitState = CircuitState.CLOSED
failures: list[float] = field(default_factory=list)
last_state_change: float = field(default_factory=time.time)
success_count: int = 0
async def execute(self, func, *args, **kwargs):
if self.state == CircuitState.OPEN:
if time.time() - self.last_state_change >= self.config.recovery_timeout:
self.state = CircuitState.HALF_OPEN
self.last_state_change = time.time()
try:
result = await func(*args, **kwargs)
self._on_success()
return result
except Exception:
self._on_failure()
raise
else:
raise CircuitOpenError("Circuit breaker is OPEN")
try:
result = await func(*args, **kwargs)
self._on_success()
return result
except Exception as e:
self._on_failure()
raise e
def _on_success(self):
self.success_count += 1
self.failures = [] # Reset failures on success in strict mode
if self.state == CircuitState.HALF_OPEN:
if self.success_count >= self.config.success_threshold:
self.state = CircuitState.CLOSED
self.last_state_change = time.time()
elif self.state == CircuitState.CLOSED:
if len(self.failures) > self.config.window_size:
self.failures.pop(0)
def _on_failure(self):
self.success_count = 0
current_time = time.time()
self.failures.append(current_time)
cutoff = current_time - 30 # Simple 30s window for example
self.failures = [t for t in self.failures if t >= cutoff]
if self.state == CircuitState.CLOSED:
recent_failures = len([f for f in self.failures if current_time - f <= 30])
if recent_failures >= self.config.failure_threshold:
self.state = CircuitState.OPEN
self.last_state_change = current_time
elif self.state == CircuitState.HALF_OPEN:
self.state = CircuitState.OPEN
self.last_state_change = current_time
class CircuitOpenError(Exception):
pass
Note that the execute method above is a wrapper. In a real agent system, you must inject semantic validation before calling _on_success or _on_failure.
For example, if the LLM returns valid JSON, it's not a "success" if the JSON contains a tool name that doesn't exist in the registry. You must parse the LLM output, validate it against your schema, and then signal success or failure to the circuit breaker. This prevents the breaker from closing too early when the LLM is "available" but "useless".
Retries in agentic systems are tricky because LLM calls are stateful. If you retry a call, you are often retrying with the same context. If the failure was due to the context being too long or confusing, a simple retry will fail.
We introduce a RetryWithContextModification pattern. Instead of blindly retrying, the orchestrator analyzes the error and modifies the prompt for the next attempt.
import asyncio
import random
class AgenticRetryPolicy:
def __init__(self, max_retries=3):
self.max_retries = max_retries
async def execute_with_retry(self, agent_step, context, error_classifier):
"""
agent_step: The function to execute (LLM call)
context: The mutable conversation/context object
error_classifier: A function that returns 'TRANSIENT' or 'PERMANENT'
"""
for attempt in range(self.max_retries):
try:
return await agent_step(context)
except Exception as e:
error_type = error_classifier(e)
if error_type == "TRANSIENT" or (error_type == "PERMANENT" and attempt < self.max_retries - 1):
base_wait = 2 ** attempt
jitter = random.uniform(0, 1)
wait_time = base_wait + jitter
if error_type == "PERMANENT" and "validation" in str(e).lower():
context.add_warning(f"Previous attempt failed validation. Error: {str(e)}. Please strictly adhere to schema.")
await asyncio.sleep(wait_time)
else:
raise e
raise Exception("Max retries exceeded")
Standard OpenTelemetry traces are insufficient for agents. You need to know why the LLM made a specific tool call. This requires Semantic Tracing.
prompt_tokens, completion_tokens, and model_id.
In addition to standard latency/error rates, expose:
Implementing a simple tracer:
from opentelemetry import trace
from opentelemetry.trace import Status, StatusCode
tracer = trace.get_tracer("agentic_system")
def trace_llm_call(model, prompt_tokens, completion_tokens, tool_calls_count):
with tracer.start_as_current_span("LLM_Call") as span:
span.set_attribute("gen_ai.system", "openai")
span.set_attribute("gen_ai.request.model", model)
span.set_attribute("gen_ai.usage.prompt_tokens", prompt_tokens)
span.set_attribute("gen_ai.usage.completion_tokens", completion_tokens)
span.set_attribute("agentic.tool_calls.count", tool_calls_count)
span.set_attribute("agentic.hallucination_detected", False)
span.set_status(Status(StatusCode.OK))
Agents can get stuck in a loop where they keep calling the same tool with the same arguments because the LLM doesn't recognize the previous failure.
SeenArguments cache within the agent's session. If the exact same tool arguments are submitted twice in a row, force a THINK step where the LLM is asked to explain Treat tokens as a hard resource limit like memory or CPU.
TokenBudgetManager. If an agent consumes 80% of its allocated budget for a sub-task, stop it and force a summary. This prevents long-running agents from burning through costs without progress.
Define a clear fallback strategy:
Agentic workflows are rarely idempotent. If the agent calls transfer_funds, you cannot safely retry it if the first call actually succeeded but the response was lost.
request_id for each tool call and ensure downstream APIs support idempotency keys. Store the Q: How do I handle LLM hallucinations that look like valid tool calls?
A: Always implement a Schema Validator layer between the LLM output and the actual tool execution. Never trust the LLM's self-report of what it intends to do. Parse the JSON, validate against the tool's Pydantic/Zod schema, and reject if it fails. If it fails, feed the validation error back to the LLM as a context warning.
Q: Should I retry on 4xx errors from the LLM provider?
A: Generally no. 400s are usually bad requests (bad schema, context too long). 429s are rate limits (retry with backoff). 401/403 are auth errors (circuit break immediately). Treat 400s as semantic failures and adjust the prompt, not just the timing.
Q: How much state should I store in the agent's memory?
A: Store decisions, not just logs. Storing the entire conversation history is expensive and confusing. Instead, store a structured state object: { current_goal, last_tool_result, pending_errors, token_budget_remaining }. This allows the LLM to
reconstruct its state without re-reading megabytes of dialogue. This is the difference between stateless retry (which can spiral) and stateful recovery (which can learn).
Naive retries fail because they retry everything. A well-designed retry policy classifies failures:
Retry-After, back off aggressively
from enum import Enum
import asyncio
from dataclasses import dataclass
class FailureType(Enum):
TRANSIENT = "transient"
RATE_LIMITED = "rate_limited"
VALIDATION = "validation"
LLM_UNCERTAIN = "llm_uncertain"
@dataclass
class RetryPolicy:
max_attempts: int = 3
base_delay: float = 1.0
max_delay: float = 30.0
async def execute_with_retry(self, coro_func, classifier):
last_error = None
for attempt in range(self.max_attempts):
try:
return await coro_func()
except Exception as e:
failure_type = classifier(e)
last_error = e
if failure_type == FailureType.VALIDATION:
raise
delay = min(
self.base_delay * (2 ** attempt),
self.max_delay
)
await asyncio.sleep(delay)
raise last_error
The key insight: the classifier function is where your domain knowledge lives. It inspects the exception and returns the right FailureType. This is your circuit breaker's input.
The circuit breaker wraps your retry policy. It has three states:
import time
from enum import Enum
class CircuitState(Enum):
CLOSED = "closed"
OPEN = "open"
HALF_OPEN = "half_open"
class CircuitBreaker:
def __init__(self, failure_threshold=5, timeout=60):
self.failure_threshold = failure_threshold
self.timeout = timeout
self.failure_count = 0
self.last_failure_time = None
self.state = CircuitState.CLOSED
async def call(self, func, *args, **kwargs):
if self.state == CircuitState.OPEN:
if time.time() - self.last_failure_time > self.timeout:
self.state = CircuitState.HALF_OPEN
else:
raise Exception("Circuit breaker is OPEN")
try:
result = await func(*args, **kwargs)
self._on_success()
return result
except Exception as e:
self._on_failure()
raise
def _on_success(self):
self.failure_count = 0
self.state = CircuitState.CLOSED
def _on_failure(self):
self.failure_count += 1
self.last_failure_time = time.time()
if self.failure_count >= self.failure_threshold:
self.state = CircuitState.OPEN
An agent without observability is a black box. You need three signals:
Log every circuit breaker state change. This tells you when your system is degrading.
For every tool call, record:
import logging
from dataclasses import dataclass
from typing import Optional
@dataclass
class DecisionTrace:
timestamp: float
goal: str
action: str
reasoning: str
result: str
confidence: float
error: Optional[str] = None
class ObservableAgent:
def __init__(self):
self.traces = []
self.logger = logging.getLogger("agent")
def record_decision(self, trace: DecisionTrace):
self.traces.append(trace)
self.logger.info(
f"Decision: {trace.action} | Confidence: {trace.confidence} | "
f"Goal: {trace.goal}"
)
if trace.error:
self.logger.error(f"Decision failed: {trace.error}")
Track token consumption per goal. If an agent is burning through tokens without progress, it's stuck in a loop.
Here's a complete agent loop that integrates circuit breaking, retry policies, and observability:
import asyncio
import json
from dataclasses import dataclass, field
from typing import List, Dict, Any
@dataclass
class AgentState:
current_goal: str
token_budget: int = 10000
pending_errors: List[str] = field(default_factory=list)
last_tool_result: str = ""
step_count: int = 0
class FaultTolerantAgent:
def __init__(self):
self.circuit_breaker = CircuitBreaker(failure_threshold=3)
self.retry_policy = RetryPolicy(max_attempts=3)
self.state = None
self.logger = logging.getLogger("agent")
def classify_error(self, error: Exception) -> FailureType:
error_str = str(error).lower()
if "rate limit" in error_str or "429" in error_str:
return FailureType.RATE_LIMITED
if "validation" in error_str or "invalid" in error_str:
return FailureType.VALIDATION
if "timeout" in error_str or "connection" in error_str:
return FailureType.TRANSIENT
return FailureType.TRANSIENT
async def execute_tool(self, tool_name: str, params: Dict[str, Any]) -> str:
async def _call():
if "fail" in tool_name:
raise Exception("Simulated tool failure")
return f"Result from {tool_name}"
return await self.retry_policy.execute_with_retry(
_call,
self.classify_error
)
async def run(self, goal: str, max_steps: int = 10):
self.state = AgentState(current_goal=goal)
for step in range(max_steps):
self.state.step_count = step
try:
tool_result = await self.circuit_breaker.call(
self.execute_tool,
"some_tool",
{"param": "value"}
)
self.state.last_tool_result = tool_result
self.logger.info(f"Step {step}: Success")
self._save_state()
except Exception as e:
error_type = self.classify_error(e)
self.state.pending_errors.append(str(e))
self.logger.error(f"Step {step} failed: {e}")
if error_type == FailureType.VALIDATION:
raise
elif self.circuit_breaker.state == CircuitState.OPEN:
self.logger.error("Circuit breaker open, stopping agent")
break
return self.state
def _save_state(self):
state_data = {
"current_goal": self.state.current_goal,
"last_tool_result": self.state.last_tool_result,
"pending_errors": self.state.pending_errors,
"token_budget_remaining": self.state.token_budget,
"step_count": self.state.step_count
}
with open(f"agent_state_{int(time.time())}.json", "w") as f:
json.dump(state_data, f, indent=2)
async def main():
agent = FaultTolerantAgent()
final_state = await agent.run("Build a weather app")
print(f"Completed {final_state.step_count} steps")
if __name__ == "__main__":
asyncio.run(main())
When an agent crashes and restarts, it should:
def load_state(self, state_file: str) -> AgentState:
with open(state_file, "r") as f:
data = json.load(f)
state = AgentState(**data)
state.pending_errors = self._validate_errors(state.pending_errors)
return state
def _validate_errors(self, errors: List[str]) -> List[str]:
valid_errors = []
for error in errors:
if self._is_error_resolved(error):
continue
valid_errors.append(error)
return valid_errors
Fault tolerance in autonomous AI agents isn't about preventing failures — it's about making failures predictable, recoverable, and informative. The three pillars work together:
The most important lesson: start simple. Begin with a basic retry loop and a state file. Add circuit breakers when you see cascading failures in production. Add sophisticated observability when you need to debug agent behavior. Over-engineering from day one creates more failure modes than it prevents.
Build the minimum viable fault tolerance, then evolve it based on real failure patterns you observe. The best systems aren't designed in isolation — they're shaped by the failures they've survived.