Complete Architecture, Design Patterns & Implementation Guide for AI Systems
TABLE OF CONTENTS
- Executive Overview
- Core Loop Architecture
- Loop Components & Subsystems
- State Management
- Control Flow & Termination Logic
- Error Handling & Recovery
- Context Management
- Verification & Validation Systems
- Cost Optimization & Token Management
- Advanced Patterns
- Implementation Examples
- Monitoring & Observability
EXECUTIVE OVERVIEW
Definition
Loop engineering is the practice of designing self-prompting autonomous systems where humans define objectives, constraints and verification rules—while the system autonomously iterates through cycles of task generation, execution, observation and refinement until goals are achieved.
Paradigm Shift
PROMPT ENGINEERING (Era 1)
User → [Write Perfect Prompt] → LLM → Answer → Copy/Paste
LOOP ENGINEERING (Era 2)
User → [Design Loop System] → [Autonomous Iterations]
├─ Generate Task
├─ Execute
├─ Verify Result
├─ Feedback
└─ Retry/Next (until goal met)
Why Loops > Prompts for Complex Tasks
| Dimension | Prompt Engineering | Loop Engineering |
|---|---|---|
| Task Length | Single pass | Iterative multi-step |
| Failure Recovery | Manual retry | Automatic with backoff |
| State Management | User memory | External persistence |
| Verification | Manual check | Automated validation |
| Tool Usage | Static composition | Dynamic, context-aware |
| Context Window | Single instance | Rolling, persistent |
| Cost Control | Unpredictable | Bounded with rules |
CORE LOOP ARCHITECTURE
1. The Canonical Loop Cycle
Every loop execution follows this deterministic sequence:
┌─────────────────────────────────────────────────────────┐
│ LOOP INITIALIZATION │
│ • Load state from storage │
│ • Validate preconditions │
│ • Inject context into LLM prompt │
└──────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ PHASE 1: PLANNING (LLM Call) │
│ │
│ Input: [Goal] + [Current State] + [Context] │
│ LLM: Generate next action(s) │
│ Output: {action, params, reasoning, confidence} │
│ │
│ Cost: ~0.5K-5K tokens depending on context size │
└──────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ PHASE 2: EXECUTION (Tool/Code Run) │
│ │
│ • Parse LLM output (JSON schema validation) │
│ • Execute tool with timeout protection │
│ • Capture stdout/stderr/errors │
│ • Record wall-clock time & resource usage │
│ │
│ Cost: $0 (deterministic execution) │
└──────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ PHASE 3: OBSERVATION (State Update) │
│ │
│ • Capture execution output │
│ • Update external state stores │
│ • Extract metrics (success/failure, delta) │
│ • Persist iteration record │
│ │
│ Cost: $0 (I/O operations) │
└──────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ PHASE 4: VERIFICATION (Goal Check) │
│ │
│ Evaluate: Goal Met? → EXIT LOOP │
│ Recoverable? → RETRY Phase 1 │
│ Dead End? → ERROR → Phase 5 │
│ New State? → LOOP again │
│ │
│ Cost: $0 (deterministic logic) │
└──────────────────┬──────────────────────────────────────┘
│
┌─────────────┴──────────┬──────────────┐
│ │ │
▼ ▼ ▼
[GOAL MET] [CONTINUE LOOP] [ERROR/TIMEOUT]
EXIT Loop back Phase 5
to Planning
2. Loop State Machine
[INIT]
│
├─→ [PLANNING] ──(LLM Call)──→ [PARSING] ──(Validate)──→ [EXECUTION]
│ │
│ ▼
│ ┌──────────────────────────────→ [OBSERVATION]
│ │ │
│ │ ▼
│ │ ┌─────────────────[VERIFICATION]
│ │ │ │
└──────────────────(Loop Control)───────┤ ├──→ [SUCCESS/EXIT]
│ │
└──(Retry Logic)─────┤
│
┌────────────────┤
│ │
[RECOVERY] [TIMEOUT/FATAL]
│ │
└────→ [CLEANUP]←─┘
│
▼
[FINAL_STATE]
3. Minimal Loop Implementation (Pseudocode)
class AutonomousLoop:
def __init__(self, goal: str, llm_client, tools: Dict[str, Callable]):
self.goal = goal
self.llm = llm_client
self.tools = tools
self.state = {"iterations": 0, "history": []}
self.max_iterations = 20
self.timeout = 300 # seconds
def run(self) -> LoopResult:
start_time = time.time()
while self.state["iterations"] < self.max_iterations:
# PHASE 1: PLANNING
prompt = self._build_prompt()
response = self.llm.complete(prompt)
action = self._parse_action(response)
# PHASE 2: EXECUTION
try:
result = self._execute_action(action, timeout=30)
except Exception as e:
result = {"error": str(e), "type": type(e).__name__}
# PHASE 3: OBSERVATION
self.state["history"].append({
"iteration": self.state["iterations"],
"action": action,
"result": result,
"timestamp": time.time()
})
# PHASE 4: VERIFICATION
verification = self._verify_goal_met(result)
if verification["goal_met"]:
return LoopResult(success=True, state=self.state)
# Check for unrecoverable conditions
if self._is_fatal_error(result) or (time.time() - start_time) > self.timeout:
return LoopResult(success=False, state=self.state, reason="fatal_error")
self.state["iterations"] += 1
return LoopResult(success=False, state=self.state, reason="max_iterations")
def _build_prompt(self) -> str:
context = self._serialize_state()
return f"""
Goal: {self.goal}
Current State:
{context}
Available Tools: {list(self.tools.keys())}
Respond ONLY in JSON format:
{{
"thought": "reasoning",
"tool": "tool_name",
"parameters": {{}},
"confidence": 0.0-1.0
}}
"""
def _execute_action(self, action: Dict, timeout: int):
tool_name = action["tool"]
tool = self.tools[tool_name]
return tool(**action["parameters"], timeout=timeout)
def _verify_goal_met(self, result: Dict) -> Dict:
# Implement domain-specific verification
return {"goal_met": result.get("status") == "success"}
LOOP COMPONENTS & SUBSYSTEMS
1. Planning Engine (LLM Interface)
Responsibilities:
- Convert current state to optimized prompt
- Call LLM with appropriate temperature/max_tokens
- Parse structured output (JSON schema validation)
- Handle token budget constraints
Implementation:
class PlanningEngine:
def __init__(self, model: str = "claude-sonnet-4-6", max_tokens: int = 2000):
self.model = model
self.max_tokens = max_tokens
self.client = Anthropic()
def generate_plan(self,
goal: str,
state: Dict,
context: str,
tools_available: List[str]) -> Dict:
"""
Generate next action(s) from LLM
Args:
goal: Loop objective
state: Current execution state
context: Historical context + observations
tools_available: Available tool names
Returns:
Parsed action dict with validation
"""
# Temperature varies by task criticality
temperature = 0.3 if state["retry_count"] > 2 else 0.7
prompt = f"""You are an autonomous AI agent executing this goal:
GOAL: {goal}
Current Execution State:
{json.dumps(state, indent=2)}
Recent Context:
{context}
Available Tools:
{', '.join(tools_available)}
You MUST respond with ONLY valid JSON (no markdown, no preamble):
{{
"thought": "brief reasoning about next step",
"action": "tool_name or 'finish'",
"parameters": {{}},
"confidence": 0.0-1.0,
"justification": "why this action progresses toward goal"
}}
"""
message = self.client.messages.create(
model=self.model,
max_tokens=self.max_tokens,
messages=[{"role": "user", "content": prompt}]
)
response_text = message.content[0].text
# Robust JSON parsing with fallback
try:
action = json.loads(response_text)
except json.JSONDecodeError:
# Try to extract JSON from markdown code blocks
match = re.search(r'```
(?:json)?\s*(.*?)\s*
```', response_text, re.DOTALL)
if match:
action = json.loads(match.group(1))
else:
raise ValueError(f"Failed to parse LLM response: {response_text}")
# Validate schema
self._validate_action_schema(action, tools_available)
return action
def _validate_action_schema(self, action: Dict, available_tools: List[str]):
required_keys = {"thought", "action", "parameters"}
if not required_keys.issubset(action.keys()):
raise ValueError(f"Missing required keys. Got: {action.keys()}")
if action["action"] not in available_tools and action["action"] != "finish":
raise ValueError(f"Unknown tool: {action['action']}")
if not isinstance(action.get("parameters"), dict):
raise ValueError("parameters must be a dict")
2. Execution Engine (Tool/Action Runner)
Responsibilities:
- Execute planned actions safely
- Manage resource limits (timeout, memory)
- Capture output and errors
- Provide structured execution feedback
Implementation:
class ExecutionEngine:
def __init__(self, tools: Dict[str, Callable], timeout: int = 30):
self.tools = tools
self.timeout = timeout
self.execution_log = []
def execute(self, action: Dict) -> ExecutionResult:
"""
Safely execute a planned action with timeout protection
"""
tool_name = action["action"]
parameters = action.get("parameters", {})
if tool_name == "finish":
return ExecutionResult(
status="success",
output={"message": "Loop completed successfully"},
tool=tool_name,
duration=0
)
if tool_name not in self.tools:
return ExecutionResult(
status="error",
error=f"Unknown tool: {tool_name}",
tool=tool_name,
duration=0
)
tool = self.tools[tool_name]
start_time = time.time()
try:
# Execute with timeout using signal or ThreadPoolExecutor
result = self._execute_with_timeout(tool, parameters)
return ExecutionResult(
status="success",
output=result,
tool=tool_name,
duration=time.time() - start_time,
parameters=parameters
)
except TimeoutError:
return ExecutionResult(
status="timeout",
error=f"Tool '{tool_name}' exceeded {self.timeout}s timeout",
tool=tool_name,
duration=time.time() - start_time
)
except Exception as e:
return ExecutionResult(
status="error",
error=str(e),
error_type=type(e).__name__,
tool=tool_name,
duration=time.time() - start_time
)
def _execute_with_timeout(self, func: Callable, params: Dict, timeout: int = None):
"""Execute function with timeout protection"""
timeout = timeout or self.timeout
from concurrent.futures import ThreadPoolExecutor, TimeoutError as FutureTimeout
with ThreadPoolExecutor(max_workers=1) as executor:
future = executor.submit(func, **params)
try:
return future.result(timeout=timeout)
except FutureTimeout:
raise TimeoutError(f"Execution exceeded {timeout}s")
3. Verification Engine (Goal & State Validation)
Responsibilities:
- Check if goal is achieved
- Validate state consistency
- Detect loops/infinite recursion
- Assess progress toward objective
Implementation:
class VerificationEngine:
def __init__(self, goal: str):
self.goal = goal
self.goal_verifier = None # Domain-specific implementation
def verify_goal_met(self, state: Dict, last_result: ExecutionResult) -> VerificationResult:
"""
Determine if the loop goal has been achieved
Returns:
VerificationResult with:
- goal_met: bool
- confidence: float [0-1]
- reason: str
- suggested_action: str (retry, continue, abort)
"""
# Check for explicit success markers
if last_result.status == "success":
if self._check_goal_success_criteria(state, last_result):
return VerificationResult(
goal_met=True,
confidence=0.95,
reason="Goal success criteria met",
suggested_action="exit"
)
# Check for fatal errors
if last_result.status == "error":
if self._is_fatal_error(last_result):
return VerificationResult(
goal_met=False,
confidence=1.0,
reason=f"Fatal error: {last_result.error}",
suggested_action="abort"
)
# Check for infinite loops (same action repeating)
if self._detect_infinite_loop(state):
return VerificationResult(
goal_met=False,
confidence=0.9,
reason="Infinite loop detected",
suggested_action="abort"
)
# Progress check - are we making forward progress?
if self._detect_no_progress(state):
return VerificationResult(
goal_met=False,
confidence=0.7,
reason="No progress for N iterations",
suggested_action="retry_different_approach"
)
# Default: continue looping
return VerificationResult(
goal_met=False,
confidence=0.5,
reason="Goal not yet achieved, continuing",
suggested_action="continue"
)
def _check_goal_success_criteria(self, state: Dict, result: ExecutionResult) -> bool:
"""
Domain-specific implementation
Examples:
- Check if output matches expected format
- Verify file exists with correct content
- Validate test suite passes
"""
# This is implemented per-domain
if self.goal_verifier:
return self.goal_verifier(state, result)
return False
def _detect_infinite_loop(self, state: Dict) -> bool:
"""Detect if same action repeats N times"""
history = state.get("history", [])
if len(history) < 3:
return False
# Check last 3 actions
last_three = [h["action"]["action"] for h in history[-3:]]
return last_three[0] == last_three[1] == last_three[2]
def _detect_no_progress(self, state: Dict) -> bool:
"""Check if state has advanced in last N iterations"""
history = state.get("history", [])
if len(history) < 5:
return False
# Measure state deltas
recent_deltas = self._compute_state_deltas(history[-5:])
return all(delta < 0.1 for delta in recent_deltas) # All changes < 10%
def _compute_state_deltas(self, history: List) -> List[float]:
"""Compute change magnitude for each iteration"""
deltas = []
for i in range(1, len(history)):
prev_state = history[i-1].get("result", {})
curr_state = history[i].get("result", {})
delta = self._state_diff_magnitude(prev_state, curr_state)
deltas.append(delta)
return deltas
def _state_diff_magnitude(self, state1: Dict, state2: Dict) -> float:
"""Compute normalized difference between states (0-1)"""
# Simple heuristic: ratio of different keys to total keys
all_keys = set(state1.keys()) | set(state2.keys())
if not all_keys:
return 0.0
different = sum(1 for k in all_keys if state1.get(k) != state2.get(k))
return different / len(all_keys)
STATE MANAGEMENT
1. State Store Architecture
Loops require external, persistent state because LLMs are stateless. State must survive loop iterations and be accessible to the planning engine.
State Layers:
┌─────────────────────────────────────────────────────┐
│ LAYER 1: WORKING STATE │
│ (In-memory, fast access) │
│ • Current iteration number │
│ • Last execution result │
│ • Short-term flags │
│ │
│ Storage: Redis, in-memory dict │
│ TTL: Duration of loop execution │
└─────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ LAYER 2: PERSISTENT STATE │
│ (Database, durable) │
│ • Full execution history │
│ • Artifacts (generated files) │
│ • Metrics & performance data │
│ │
│ Storage: PostgreSQL, S3, Git │
│ TTL: Indefinite │
└─────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ LAYER 3: CONTEXT MEMORY │
│ (Long-form reference for LLM) │
│ • Previous iterations summaries │
│ • Relevant code/docs │
│ • Error patterns & solutions │
│ │
│ Storage: Vector DB, File System │
│ TTL: Loop-scoped or permanent │
└─────────────────────────────────────────────────────┘
2. State Schema
@dataclass
class LoopState:
"""Complete state snapshot for a loop execution"""
# Metadata
loop_id: str # UUID
goal: str # Loop objective
created_at: float # Unix timestamp
# Execution State
iterations: int # Current iteration count
max_iterations: int # Termination limit
status: str # "running", "success", "failed", "timeout"
# History
history: List[IterationRecord] # All iteration results
# Artifacts
artifacts: Dict[str, str] # {"file_path": "content", ...}
generated_files: List[str] # Files created during loop
# Metrics
total_tokens_used: int # Sum of all LLM tokens
total_cost: float # USD cost estimate
wall_clock_duration: float # Seconds
# Context
context_window: str # Serialized state for next LLM call
error_log: List[str] # Error messages for recovery
def serialize(self) -> str:
"""Serialize to JSON for storage"""
return json.dumps(asdict(self), default=str)
@classmethod
def deserialize(cls, data: str) -> 'LoopState':
"""Load from JSON"""
return cls(**json.loads(data))
def to_context_string(self) -> str:
"""Serialize state for LLM prompt injection"""
return f"""
EXECUTION STATE:
- Iteration: {self.iterations}/{self.max_iterations}
- Status: {self.status}
- Tokens Used: {self.total_tokens_used}
- Cost: ${self.total_cost:.4f}
- Duration: {self.wall_clock_duration:.1f}s
RECENT HISTORY:
{self._serialize_recent_history()}
ARTIFACTS CREATED:
{json.dumps(self.generated_files, indent=2)}
ERROR LOG:
{json.dumps(self.error_log[-5:], indent=2)} # Last 5 errors
"""
def _serialize_recent_history(self, limit: int = 5) -> str:
recent = self.history[-limit:]
return "\n".join([
f" [{r.iteration}] {r.action['action']} → {r.result['status']}"
for r in recent
])
@dataclass
class IterationRecord:
"""Single iteration record"""
iteration: int
timestamp: float
action: Dict # {"action": "...", "parameters": {...}}
result: ExecutionResult
tokens_used: int
cost: float
duration: float
3. State Persistence
class StateStore:
"""Unified state persistence layer"""
def __init__(self, backend: str = "postgres"):
self.backend = backend
if backend == "postgres":
self.db = PostgresBackend()
elif backend == "redis":
self.db = RedisBackend()
elif backend == "sqlite":
self.db = SQLiteBackend()
def save_state(self, state: LoopState) -> None:
"""Persist state after each iteration"""
self.db.insert("loop_states", {
"loop_id": state.loop_id,
"goal": state.goal,
"iteration": state.iterations,
"status": state.status,
"state_json": state.serialize(),
"updated_at": datetime.utcnow().isoformat(),
"tokens_used": state.total_tokens_used,
"cost": state.total_cost
})
def load_state(self, loop_id: str) -> LoopState:
"""Restore state from storage"""
record = self.db.query_one(
"SELECT state_json FROM loop_states WHERE loop_id = %s ORDER BY updated_at DESC",
(loop_id,)
)
return LoopState.deserialize(record["state_json"])
def save_iteration(self, loop_id: str, record: IterationRecord) -> None:
"""Append iteration record"""
self.db.insert("iteration_records", {
"loop_id": loop_id,
"iteration_num": record.iteration,
"action": json.dumps(record.action),
"result": json.dumps(asdict(record.result)),
"tokens_used": record.tokens_used,
"cost": record.cost,
"duration": record.duration,
"created_at": datetime.fromtimestamp(record.timestamp).isoformat()
})
def get_history(self, loop_id: str, limit: int = 20) -> List[IterationRecord]:
"""Fetch iteration history"""
records = self.db.query(
"""SELECT * FROM iteration_records
WHERE loop_id = %s
ORDER BY iteration_num DESC
LIMIT %s""",
(loop_id, limit)
)
return [IterationRecord(**r) for r in records]
CONTROL FLOW & TERMINATION LOGIC
1. Termination Conditions
A loop must have multiple, explicit exit criteria:
class TerminationController:
"""Manages loop exit conditions"""
def __init__(self,
max_iterations: int = 20,
timeout_seconds: int = 300,
max_cost_usd: float = 10.0):
self.max_iterations = max_iterations
self.timeout = timeout_seconds
self.max_cost = max_cost_usd
self.iteration_timeout = 30 # Per-iteration limit
def should_continue(self, state: LoopState) -> Tuple[bool, str]:
"""
Determine if loop should continue
Returns: (should_continue: bool, reason: str)
"""
# Check 1: Goal achieved
if state.status == "success":
return (False, "Goal achieved")
# Check 2: Max iterations exceeded
if state.iterations >= self.max_iterations:
return (False, f"Max iterations ({self.max_iterations}) reached")
# Check 3: Total timeout exceeded
if state.wall_clock_duration > self.timeout:
return (False, f"Total timeout ({self.timeout}s) exceeded")
# Check 4: Cost limit exceeded
if state.total_cost >= self.max_cost:
return (False, f"Cost limit (${self.max_cost}) exceeded")
# Check 5: Fatal error occurred
if state.status == "fatal_error":
return (False, "Fatal error encountered")
# Check 6: Unrecoverable state
if self._is_unrecoverable(state):
return (False, "Unrecoverable state detected")
# Otherwise, continue
return (True, "Continue looping")
def _is_unrecoverable(self, state: LoopState) -> bool:
"""Check for states that cannot recover"""
# Pattern: Same error repeating N times
if len(state.error_log) >= 3:
recent_errors = state.error_log[-3:]
if all(e == recent_errors[0] for e in recent_errors):
return True
# Pattern: Cycling between two states infinitely
if len(state.history) >= 4:
actions = [h.action["action"] for h in state.history[-4:]]
if actions[0] == actions[2] and actions[1] == actions[3]:
return True
return False
2. Retry & Backoff Strategy
class RetryController:
"""Intelligent retry logic with exponential backoff"""
def __init__(self,
max_retries: int = 3,
backoff_base: float = 2.0,
jitter: bool = True):
self.max_retries = max_retries
self.backoff_base = backoff_base
self.jitter = jitter
self.retry_counts = {} # {action_id: count}
def should_retry(self, state: LoopState, last_result: ExecutionResult) -> bool:
"""Determine if action should be retried"""
# Never retry if goal is achieved
if state.status == "success":
return False
# Only retry on recoverable errors
if last_result.status not in ["timeout", "error"]:
return False
# Check retry limit
action_id = state.history[-1].action.get("action")
retries = self.retry_counts.get(action_id, 0)
if retries >= self.max_retries:
return False
# Don't retry fatal errors
if self._is_fatal_error(last_result):
return False
return True
def get_backoff_delay(self, retry_count: int) -> float:
"""Calculate exponential backoff with optional jitter"""
delay = self.backoff_base ** retry_count
if self.jitter:
delay *= (0.5 + random.random()) # 50-150% of computed delay
return delay
def _is_fatal_error(self, result: ExecutionResult) -> bool:
"""Identify errors that shouldn't be retried"""
fatal_errors = [
"KeyError", "ValueError", "AttributeError", # Code bugs
"PermissionError", "FileNotFoundError", # Resource errors
"SyntaxError", "ImportError" # Config errors
]
return result.error_type in fatal_errors
ERROR HANDLING & RECOVERY
1. Error Classification
class ErrorClassification:
"""Categorize errors for appropriate handling"""
TRANSIENT = {
"timeout": "Exceeded execution time limit",
"rate_limit": "API rate limit exceeded",
"network": "Network connectivity issue",
"memory": "Temporary memory pressure"
}
RECOVERABLE = {
"invalid_tool": "Tool doesn't exist, try different tool",
"malformed_params": "Parameter format incorrect, fix and retry",
"missing_dependency": "Required file/tool missing, create it first"
}
FATAL = {
"permission_denied": "Insufficient permissions, cannot proceed",
"syntax_error": "Code has syntax errors, manual fix required",
"resource_exhausted": "Cannot allocate required resources",
"auth_failed": "Authentication failure, credentials invalid"
}
@classmethod
def classify(cls, error: Exception, context: Dict) -> Tuple[str, str]:
"""
Classify error and suggest recovery action
Returns: (category, suggested_action)
"""
error_name = type(error).__name__
error_msg = str(error).lower()
# Match against patterns
for category, errors in [
("transient", cls.TRANSIENT),
("recoverable", cls.RECOVERABLE),
("fatal", cls.FATAL)
]:
for error_key, description in errors.items():
if error_key in error_msg or error_key in error_name.lower():
return (category, error_key)
# Default classification
if "timeout" in error_msg:
return ("transient", "timeout")
if "error" in error_msg:
return ("recoverable", "unknown_error")
return ("fatal", "unclassified")
class ErrorRecoveryStrategy:
"""Map error classifications to recovery actions"""
RECOVERY_ACTIONS = {
"timeout": {
"action": "retry_with_longer_timeout",
"max_retries": 2,
"delay_seconds": 5
},
"rate_limit": {
"action": "backoff_and_retry",
"max_retries": 3,
"delay_seconds": 30
},
"invalid_tool": {
"action": "replan_with_available_tools",
"suggest_alternatives": True,
"max_retries": 1
},
"missing_dependency": {
"action": "create_dependency_first",
"max_retries": 2
},
"syntax_error": {
"action": "abort_loop",
"log_error": True
}
}
@staticmethod
def handle_error(error: Exception,
error_type: str,
state: LoopState) -> Dict:
"""
Handle error and return recovery strategy
"""
strategy = ErrorRecoveryStrategy.RECOVERY_ACTIONS.get(
error_type,
{"action": "abort_loop"}
)
return {
"should_retry": strategy.get("max_retries", 0) > 0,
"action": strategy["action"],
"delay": strategy.get("delay_seconds", 0),
"retry_limit": strategy.get("max_retries", 0),
"log_context": {
"error": str(error),
"iteration": state.iterations,
"cost_so_far": state.total_cost
}
}
2. Circuit Breaker Pattern
class CircuitBreaker:
"""Prevent cascading failures with circuit breaker pattern"""
def __init__(self, failure_threshold: int = 5, timeout: int = 60):
self.failure_threshold = failure_threshold
self.timeout = timeout
self.failure_count = 0
self.last_failure_time = None
self.state = "closed" # "closed", "open", "half_open"
def record_success(self):
"""Mark successful execution"""
self.failure_count = 0
self.state = "closed"
def record_failure(self):
"""Mark failed execution"""
self.failure_count += 1
self.last_failure_time = time.time()
if self.failure_count >= self.failure_threshold:
self.state = "open"
raise CircuitBreakerOpen(
f"Circuit opened after {self.failure_threshold} failures"
)
def can_execute(self) -> bool:
"""Check if action should be attempted"""
if self.state == "closed":
return True
if self.state == "open":
# Try to transition to half-open if timeout passed
if time.time() - self.last_failure_time > self.timeout:
self.state = "half_open"
return True
return False
if self.state == "half_open":
return True
return False
def on_half_open_success(self):
"""Reset after successful test in half-open state"""
self.failure_count = 0
self.state = "closed"
def on_half_open_failure(self):
"""Return to open state"""
self.state = "open"
self.last_failure_time = time.time()
CONTEXT MANAGEMENT
1. Rolling Context Window
Since LLMs have finite context windows, loops must manage what information flows to the planning engine:
class ContextManager:
"""Manage context injection for LLM planning"""
def __init__(self, max_context_tokens: int = 8000):
self.max_context_tokens = max_context_tokens
self.tokenizer = Tokenizer() # Claude tokenizer
def build_planning_context(self,
state: LoopState,
goal: str,
tools: List[str]) -> str:
"""
Build optimized context for planning LLM call
Strategy:
1. Goal (fixed)
2. Current iteration state (variable)
3. Recent history summaries (compressed)
4. Relevant artifacts (sampled)
5. Error context (if applicable)
"""
sections = []
tokens_used = 0
# SECTION 1: Goal (Always included, high priority)
goal_section = f"OBJECTIVE:\n{goal}\n\n"
sections.append(("goal", goal_section))
tokens_used += self.tokenizer.count_tokens(goal_section)
# SECTION 2: Current State
state_section = f"""CURRENT STATE (Iteration {state.iterations}/{state.max_iterations}):
- Status: {state.status}
- Cost So Far: ${state.total_cost:.4f}
- Duration: {state.wall_clock_duration:.1f}s
- Available Tools: {', '.join(tools)}
"""
sections.append(("state", state_section))
tokens_used += self.tokenizer.count_tokens(state_section)
# SECTION 3: Recent History (Compressed)
history_section = self._compress_history(state.history[-10:])
sections.append(("history", history_section))
tokens_used += self.tokenizer.count_tokens(history_section)
# SECTION 4: Artifacts (Sampled if too large)
if tokens_used < self.max_context_tokens * 0.7:
artifacts_section = self._sample_artifacts(state.artifacts)
sections.append(("artifacts", artifacts_section))
tokens_used += self.tokenizer.count_tokens(artifacts_section)
# SECTION 5: Error Context (If errors occurred)
if state.error_log:
errors_section = f"""RECENT ERRORS:
{chr(10).join(state.error_log[-3:])}
"""
sections.append(("errors", errors_section))
tokens_used += self.tokenizer.count_tokens(errors_section)
# Trim if over budget
if tokens_used > self.max_context_tokens:
sections = self._trim_sections(sections, self.max_context_tokens)
return "\n".join([section for _, section in sections])
def _compress_history(self, history: List[IterationRecord]) -> str:
"""Summarize history in compact format"""
lines = ["HISTORY:"]
for record in history[-10:]:
action = record.action.get("action", "unknown")
status = record.result.get("status", "unknown")
lines.append(f" [{record.iteration}] {action} → {status}")
return "\n".join(lines) + "\n"
def _sample_artifacts(self, artifacts: Dict[str, str]) -> str:
"""Include relevant artifacts, truncate if needed"""
section = "GENERATED ARTIFACTS:\n"
for path, content in artifacts.items():
# Include first 500 chars
preview = content[:500] + ("..." if len(content) > 500 else "")
section += f"\n{path}:\n{preview}\n"
return section
def _trim_sections(self, sections: List[Tuple], max_tokens: int) -> List[Tuple]:
"""Drop least important sections to fit budget"""
priority = ["goal", "state", "history", "artifacts", "errors"]
sorted_sections = sorted(
sections,
key=lambda x: priority.index(x[0]) if x[0] in priority else 999
)
trimmed = []
tokens = 0
for name, section in sorted_sections:
section_tokens = self.tokenizer.count_tokens(section)
if tokens + section_tokens <= max_tokens:
trimmed.append((name, section))
tokens += section_tokens
return trimmed
2. Memory Types in Loops
class LoopMemory:
"""Multi-tier memory system for loops"""
def __init__(self):
self.short_term = {} # Iteration-level state
self.working = {} # Task-level artifacts
self.long_term = {} # Loop-lifetime knowledge
self.external = {} # Files, Git, databases
# SHORT-TERM (lasts 1-2 iterations)
# Used for: immediate action results, flags
# WORKING (lasts duration of loop)
# Used for: generated code, intermediate outputs
# LONG-TERM (persists across loops)
# Used for: learned patterns, solutions to recurring problems
# EXTERNAL (persists indefinitely)
# Used for: final artifacts, version control, audit log
VERIFICATION & VALIDATION SYSTEMS
1. Multi-Layer Verification
class VerificationFramework:
"""Multi-layer verification for robust goal achievement"""
def __init__(self):
self.assertions = []
self.validators = []
def register_assertion(self, name: str, predicate: Callable[[Dict], bool]):
"""Register a verifiable assertion about success"""
self.assertions.append({"name": name, "check": predicate})
def register_validator(self, name: str, validator: Callable[[Dict], bool]):
"""Register a validator that must pass"""
self.validators.append({"name": name, "check": validator})
def verify_goal_achieved(self, state: LoopState, result: Dict) -> VerificationReport:
"""
Multi-layer verification:
Layer 1: Assertions (conditions that must be true)
Layer 2: Validators (functional validation)
Layer 3: Integration tests (end-to-end)
"""
report = VerificationReport()
# LAYER 1: Assertions
assertion_results = []
for assertion in self.assertions:
try:
passed = assertion["check"](result)
assertion_results.append({
"assertion": assertion["name"],
"passed": passed,
"weight": 1.0
})
except Exception as e:
assertion_results.append({
"assertion": assertion["name"],
"passed": False,
"error": str(e)
})
# LAYER 2: Validators
validator_results = []
for validator in self.validators:
try:
passed = validator["check"](result)
validator_results.append({
"validator": validator["name"],
"passed": passed,
"required": True
})
except Exception as e:
validator_results.append({
"validator": validator["name"],
"passed": False,
"error": str(e)
})
# LAYER 3: Integration Check
integration_passed = self._run_integration_test(result)
# Aggregate
all_assertions_pass = all(a["passed"] for a in assertion_results)
all_validators_pass = all(v["passed"] for v in validator_results)
report.success = all_assertions_pass and all_validators_pass and integration_passed
report.confidence = self._compute_confidence(
assertion_results, validator_results, integration_passed
)
report.details = {
"assertions": assertion_results,
"validators": validator_results,
"integration": integration_passed
}
return report
def _run_integration_test(self, result: Dict) -> bool:
"""Run end-to-end integration test"""
# Domain-specific implementation
try:
# Example: Run test suite, verify artifact quality, etc.
return True
except:
return False
def _compute_confidence(self, assertions, validators, integration) -> float:
"""Compute confidence [0-1]"""
scores = []
# Assertion confidence
assertion_pass_rate = sum(1 for a in assertions if a["passed"]) / len(assertions)
scores.append(assertion_pass_rate * 0.4)
# Validator confidence
validator_pass_rate = sum(1 for v in validators if v["passed"]) / len(validators)
scores.append(validator_pass_rate * 0.4)
# Integration confidence
scores.append(integration * 0.2)
return sum(scores)
@dataclass
class VerificationReport:
success: bool
confidence: float # [0-1]
details: Dict
COST OPTIMIZATION & TOKEN MANAGEMENT
1. Token Budget System
class TokenBudgetManager:
"""Track and optimize token usage across loop"""
def __init__(self, budget_usd: float = 10.0, model: str = "claude-sonnet-4-6"):
self.budget_usd = budget_usd
self.model = model
self.pricing = self._load_pricing()
self.spent = 0.0
self.token_log = []
def _load_pricing(self) -> Dict:
"""Current pricing (June 2026)"""
return {
"claude-opus-4-6": {"input": 0.015/1000, "output": 0.045/1000},
"claude-sonnet-4-6": {"input": 0.003/1000, "output": 0.015/1000},
"claude-haiku-4-5": {"input": 0.0008/1000, "output": 0.004/1000}
}
def can_afford_call(self, estimated_input_tokens: int) -> bool:
"""Check if next LLM call fits budget"""
estimated_cost = self._estimate_cost(estimated_input_tokens, 2000)
return self.spent + estimated_cost <= self.budget_usd
def record_call(self, input_tokens: int, output_tokens: int) -> float:
"""Log token usage and return cost"""
cost = self._compute_cost(input_tokens, output_tokens)
self.spent += cost
self.token_log.append({
"input": input_tokens,
"output": output_tokens,
"cost": cost,
"timestamp": time.time()
})
return cost
def _estimate_cost(self, input_tokens: int, output_tokens: int) -> float:
"""Estimate cost before call"""
pricing = self.pricing[self.model]
return (input_tokens * pricing["input"] +
output_tokens * pricing["output"])
def _compute_cost(self, input_tokens: int, output_tokens: int) -> float:
"""Compute actual cost"""
return self._estimate_cost(input_tokens, output_tokens)
def get_remaining_budget(self) -> float:
return self.budget_usd - self.spent
def get_budget_report(self) -> Dict:
return {
"total_budget": self.budget_usd,
"spent": self.spent,
"remaining": self.get_remaining_budget(),
"percent_used": (self.spent / self.budget_usd) * 100,
"total_tokens": sum(t["input"] + t["output"] for t in self.token_log),
"call_count": len(self.token_log)
}
class TokenOptimizer:
"""Optimize token usage through compression and caching"""
@staticmethod
def compress_history(history: List[IterationRecord], ratio: float = 0.5) -> str:
"""Compress history summaries to reduce context size"""
# Keep first and last iterations completely
# Summarize middle iterations
if len(history) <= 4:
return "\n".join(f"[{r.iteration}] {r.action['action']}" for r in history)
compressed = []
compressed.append(f"[{history[0].iteration}] {history[0].action['action']}")
# Compress middle
middle = history[1:-1]
summary = f"... {len(middle)} intermediate iterations (actions: {set(m.action['action'] for m in middle)}) ..."
compressed.append(summary)
compressed.append(f"[{history[-1].iteration}] {history[-1].action['action']}")
return "\n".join(compressed)
@staticmethod
def cache_results(result: Dict, key: str) -> None:
"""Cache successful results to avoid recomputation"""
import hashlib
hash_key = hashlib.md5(json.dumps(result).encode()).hexdigest()
# Store in Redis or cache system
pass
ADVANCED PATTERNS
1. Parallel Loop Execution
class ParallelLoopCoordinator:
"""Coordinate multiple parallel loops with shared state"""
def __init__(self, max_parallel: int = 3):
self.max_parallel = max_parallel
self.loops = {}
self.shared_state = {}
self.lock = threading.Lock()
def spawn_loop(self, loop_id: str, goal: str) -> 'AutonomousLoop':
"""Create new autonomous loop"""
if len(self.loops) >= self.max_parallel:
raise RuntimeError("Max parallel loops reached")
loop = AutonomousLoop(goal)
self.loops[loop_id] = loop
return loop
def run_parallel(self) -> Dict[str, LoopResult]:
"""Execute all loops in parallel"""
futures = {}
with ThreadPoolExecutor(max_workers=self.max_parallel) as executor:
for loop_id, loop in self.loops.items():
futures[loop_id] = executor.submit(loop.run)
results = {}
for loop_id, future in futures.items():
results[loop_id] = future.result()
return results
def coordinate_state(self, loop_id: str, state_update: Dict) -> None:
"""Share state between parallel loops"""
with self.lock:
self.shared_state.update(state_update)
2. Nested Loops
class NestedLoopController:
"""Support hierarchical/nested loop execution"""
def __init__(self, parent_goal: str):
self.parent_goal = parent_goal
self.subloops = []
self.parent_state = {}
def create_subloop(self, subgoal: str) -> 'AutonomousLoop':
"""Create nested loop for subtask"""
subloop = AutonomousLoop(subgoal)
subloop.parent_state = self.parent_state
self.subloops.append(subloop)
return subloop
def run_hierarchical(self) -> LoopResult:
"""Execute parent loop with subloop management"""
for subloop in self.subloops:
result = subloop.run()
if not result.success:
return LoopResult(success=False, reason="subloop_failed")
return LoopResult(success=True)
3. Adaptive Loop Configuration
class AdaptiveLoopController:
"""Adjust loop parameters based on performance"""
def __init__(self, initial_config: Dict):
self.config = initial_config
self.metrics = []
def adapt_parameters(self, state: LoopState) -> None:
"""Dynamically adjust loop behavior based on metrics"""
# Measure success rate
success_rate = self._compute_success_rate(state)
# Adjust temperature based on performance
if success_rate < 0.3:
# Increase determinism for struggling loops
self.config["temperature"] = 0.3
elif success_rate > 0.8:
# Can afford more exploration
self.config["temperature"] = 0.8
# Adjust max_iterations based on progress
if self._is_making_progress(state):
self.config["max_iterations"] = min(
self.config["max_iterations"] + 5, 50
)
else:
self.config["max_iterations"] = max(
self.config["max_iterations"] - 5, 10
)
IMPLEMENTATION EXAMPLES
Example 1: Code Generation Loop
def create_code_generation_loop(
requirements: str,
language: str = "python",
budget: float = 5.0
) -> AutonomousLoop:
"""Create a loop that generates and tests code"""
def generate_tool(specs: str) -> Dict:
"""Generate code based on specs"""
# Implementation: Call LLM to generate code
pass
def test_tool(code: str) -> Dict:
"""Execute generated code and test it"""
try:
exec_globals = {}
exec(code, exec_globals)
return {"status": "success", "output": exec_globals}
except Exception as e:
return {"status": "error", "error": str(e)}
def lint_tool(code: str) -> Dict:
"""Lint and validate code quality"""
# Run pylint, mypy, etc.
pass
loop = AutonomousLoop(
goal=requirements,
llm_client=Anthropic(),
tools={
"generate": generate_tool,
"test": test_tool,
"lint": lint_tool
}
)
loop.budget_manager = TokenBudgetManager(budget)
return loop
Example 2: Data Processing Loop
def create_data_processing_loop(
input_file: str,
processing_spec: str
) -> AutonomousLoop:
"""Create a loop that processes data iteratively"""
def load_data_tool(path: str) -> Dict:
import pandas as pd
return {"data": pd.read_csv(path).to_dict()}
def transform_tool(data: Dict, transformation: str) -> Dict:
# Apply transformation
pass
def validate_tool(data: Dict) -> Dict:
# Validate data quality
pass
def save_tool(data: Dict, path: str) -> Dict:
import pandas as pd
pd.DataFrame(data).to_csv(path)
return {"status": "success", "path": path}
loop = AutonomousLoop(
goal=processing_spec,
tools={
"load": load_data_tool,
"transform": transform_tool,
"validate": validate_tool,
"save": save_tool
}
)
return loop
MONITORING & OBSERVABILITY
1. Loop Metrics & Telemetry
class LoopTelemetry:
"""Comprehensive monitoring for loop execution"""
def __init__(self, loop_id: str):
self.loop_id = loop_id
self.metrics = {
"iterations": 0,
"successes": 0,
"failures": 0,
"retries": 0,
"total_tokens": 0,
"total_cost": 0.0,
"start_time": time.time(),
"errors": []
}
def record_iteration(self, success: bool, tokens: int, cost: float) -> None:
self.metrics["iterations"] += 1
if success:
self.metrics["successes"] += 1
else:
self.metrics["failures"] += 1
self.metrics["total_tokens"] += tokens
self.metrics["total_cost"] += cost
def record_error(self, error: Exception) -> None:
self.metrics["errors"].append({
"type": type(error).__name__,
"message": str(error),
"timestamp": time.time()
})
def get_metrics(self) -> Dict:
duration = time.time() - self.metrics["start_time"]
return {
**self.metrics,
"duration_seconds": duration,
"success_rate": self.metrics["successes"] / max(self.metrics["iterations"], 1),
"cost_per_iteration": self.metrics["total_cost"] / max(self.metrics["iterations"], 1),
"tokens_per_iteration": self.metrics["total_tokens"] / max(self.metrics["iterations"], 1)
}
class LoopDashboard:
"""Real-time dashboard for loop monitoring"""
def __init__(self):
self.loops = {}
def register_loop(self, loop_id: str, telemetry: LoopTelemetry):
self.loops[loop_id] = telemetry
def render_dashboard(self) -> str:
"""Generate text-based dashboard"""
output = "LOOP EXECUTION DASHBOARD\n"
output += "=" * 80 + "\n\n"
for loop_id, telemetry in self.loops.items():
metrics = telemetry.get_metrics()
output += f"Loop: {loop_id}\n"
output += f" Iterations: {metrics['iterations']}\n"
output += f" Success Rate: {metrics['success_rate']:.1%}\n"
output += f" Total Cost: ${metrics['total_cost']:.4f}\n"
output += f" Duration: {metrics['duration_seconds']:.1f}s\n"
output += f" Tokens/Iteration: {metrics['tokens_per_iteration']:.0f}\n"
output += "\n"
return output
2. Logging & Audit Trail
class LoopLogger:
"""Comprehensive audit logging for compliance and debugging"""
def __init__(self, loop_id: str, log_file: str = None):
self.loop_id = loop_id
self.logger = self._setup_logger(log_file)
def _setup_logger(self, log_file: str):
import logging
logger = logging.getLogger(f"loop.{self.loop_id}")
formatter = logging.Formatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
# File handler
if log_file:
fh = logging.FileHandler(log_file)
fh.setLevel(logging.DEBUG)
fh.setFormatter(formatter)
logger.addHandler(fh)
# Console handler
ch = logging.StreamHandler()
ch.setLevel(logging.INFO)
ch.setFormatter(formatter)
logger.addHandler(ch)
return logger
def log_iteration(self, iteration: int, action: Dict, result: Dict) -> None:
self.logger.info(
f"[ITERATION {iteration}] Action: {action['action']} → "
f"Status: {result.get('status')}"
)
def log_error(self, error: Exception, context: Dict) -> None:
self.logger.error(
f"ERROR: {type(error).__name__}: {str(error)}\n"
f"Context: {json.dumps(context, indent=2)}"
)
def log_state_snapshot(self, state: LoopState) -> None:
self.logger.debug(
f"STATE SNAPSHOT:\n{json.dumps(json.loads(state.serialize()), indent=2)}"
)
PRODUCTION DEPLOYMENT CHECKLIST
- [ ] Token budget enforcement implemented
- [ ] Circuit breaker protection active
- [ ] Cost limits enforced per-loop
- [ ] Comprehensive error classification & recovery
- [ ] State persistence configured (DB backend)
- [ ] Monitoring/telemetry in place
- [ ] Retry logic with exponential backoff
- [ ] Verification framework implemented
- [ ] Context management optimized
- [ ] Logging/audit trail enabled
- [ ] Graceful degradation strategies
- [ ] Performance benchmarking completed
- [ ] Cost projections validated
- [ ] Runaway loop detection active
CONCLUSION
Loop engineering shifts AI development from reactive prompting to proactive system design. The key is building robust infrastructure around iterative autonomy: state management, verification, cost control and error recovery.
The loop becomes the unit of work, not the LLM call.
Top comments (1)
Reading the two loop detectors next to each other, one gap jumped out. VerificationEngine._detect_infinite_loop only fires when the same action repeats three times straight, so a loop that oscillates between two actions (A, B, A, B) slides right past it. Your TerminationController._is_unrecoverable already has the two-cycle check for exactly that case (actions[0] == actions[2] and actions[1] == actions[3]), so the fix is just sharing that logic rather than running the weaker check in one place and the stronger one in the other. Minor, but A/B oscillation is one of the more common ways these loops actually get stuck.