nous-hermeshub Autonomous Agents
Autonomous agents are AI systems that can independently decompose goals, plan actions, execute tools, and self-correct without constant human guidance. The challenge isn't making them capable - it's making them reliable. Every extra decision multiplies failure probability.
This skill covers agent loops (ReAct, Plan-Execute), goal decomposition, reflection patterns, and production reliability. Key insight: compounding error rates kill autonomous agents. A 95% success rate per step drops to 60% by step 10. Build for reliability first, autonomy second.
2025 lesson: The winners are constrained, domain-specific agents with clear boundaries, not "autonomous everything." Treat AI outputs as proposals, not truth.
Principles
- Reliability over autonomy - every step compounds error probability
- Constrain scope - domain-specific beats general-purpose
- Treat outputs as proposals, not truth
- Build guardrails before expanding capabilities
- Human-in-the-loop for critical decisions is non-negotiable
- Log everything - every action must be auditable
- Fail safely with rollback, not silently with corruption
Capabilities
- autonomous-agents
- agent-loops
- goal-decomposition
- self-correction
- reflection-patterns
- react-pattern
- plan-execute
- agent-reliability
- agent-guardrails
Scope
- multi-agent-systems → multi-agent-orchestration
- tool-building → agent-tool-builder
- memory-systems → agent-memory-systems
- workflow-orchestration → workflow-automation
Tooling
Frameworks
- LangGraph - When: Production agents with state management Note: 1.0 released Oct 2025, checkpointing, human-in-loop
- AutoGPT - When: Research/experimentation, open-ended exploration Note: Needs external guardrails for production
- CrewAI - When: Role-based agent teams Note: Good for specialized agent collaboration
- Claude Agent SDK - When: Anthropic ecosystem agents Note: Computer use, tool execution
Patterns
- ReAct - When: Reasoning + Acting in alternating steps Note: Foundation for most modern agents
- Plan-Execute - When: Separate planning from execution Note: Better for complex multi-step tasks
- Reflection - When: Self-evaluation and correction Note: Evaluator-optimizer loop
Patterns
ReAct Agent Loop
Alternating reasoning and action steps
When to use: Interactive problem-solving, tool use, exploration
REACT PATTERN:
""" The ReAct loop:
- Thought: Reason about what to do next
- Action: Choose and execute a tool
- Observation: Receive result
- Repeat until goal achieved
Key: Explicit reasoning traces make debugging possible """
Basic ReAct Implementation
""" from langchain.agents import create_react_agent from langchain_openai import ChatOpenAI
Define the ReAct prompt template
react_prompt = ''' Answer the question using the following format:
Question: the input question Thought: reason about what to do Action: tool_name Action Input: input to the tool Observation: result of the action ... (repeat Thought/Action/Observation as needed) Thought: I now know the final answer Final Answer: the answer '''
Create the agent
agent = create_react_agent( llm=ChatOpenAI(model="gpt-4o"), tools=tools, prompt=react_prompt, )
Execute with step limit
result = agent.invoke( {"input": query}, config={"max_iterations": 10} # Prevent runaway loops ) """
LangGraph ReAct (Production)
""" from langgraph.prebuilt import create_react_agent from langgraph.checkpoint.postgres import PostgresSaver
Production checkpointer
checkpointer = PostgresSaver.from_conn_string( os.environ["POSTGRES_URL"] )
agent = create_react_agent( model=llm, tools=tools, checkpointer=checkpointer, # Durable state )
Invoke with thread for state persistence
config = {"configurable": {"thread_id": "user-123"}} result = agent.invoke({"messages": [query]}, config) """
Plan-Execute Pattern
Separate planning phase from execution
When to use: Complex multi-step tasks, when full plan visibility matters
PLAN-EXECUTE PATTERN:
""" Two-phase approach:
- Planning: Decompose goal into subtasks
- Execution: Execute subtasks, potentially re-plan
Advantages:
- Full visibility into plan before execution
- Can validate/modify plan with human
- Cleaner separation of concerns
Disadvantages:
- Less adaptive to mid-task discoveries
- Plan may become stale """
LangGraph Plan-Execute
""" from langgraph.prebuilt import create_plan_and_execute_agent
Planner creates the task list
planner_prompt = ''' For the given objective, create a step-by-step plan. Each step should be atomic and actionable. Format: numbered list of steps. '''
Executor handles individual steps
executor_prompt = ''' You are executing step {step_number} of the plan. Previous results: {previous_results} Current step: {current_step} Execute this step using available tools. '''
agent = create_plan_and_execute_agent( planner=planner_llm, executor=executor_llm, tools=tools, replan_on_error=True, # Re-plan if step fails )
Human approval of plan
config = { "configurable": { "thread_id": "task-456", }, "interrupt_before": ["execute"], # Pause before execution }
First call creates plan
plan = agent.invoke({"objective": goal}, config)
Review plan, then continue
if human_approves(plan): result = agent.invoke(None, config) # Continue from checkpoint """
Decomposition Strategies
"""
Decomposition-First: Plan everything, then execute
Best for: Stable tasks, need full plan approval
Interleaved: Plan one step, execute, repeat
Best for: Dynamic tasks, learning as you go
def interleaved_execute(goal, max_steps=10): state = {"goal": goal, "completed": [], "remaining": [goal]}
for step in range(max_steps):
# Plan next action based on current state
next_action = planner.plan_next(state)
if next_action == "DONE":
break
# Execute and update state
result = executor.execute(next_action)
state["completed"].append((next_action, result))
# Re-evaluate remaining work
state["remaining"] = planner.reassess(state)
return state
"""
Reflection Pattern
Self-evaluation and iterative improvement
When to use: Quality matters, complex outputs, creative tasks
REFLECTION PATTERN:
""" Self-correction loop:
- Generate initial output
- Evaluate against criteria
- Critique and identify issues
- Refine based on critique
- Repeat until satisfactory
Also called: Evaluator-Optimizer, Self-Critique """
Basic Reflection
""" def reflect_and_improve(task, max_iterations=3): # Initial generation output = generator.generate(task)
for i in range(max_iterations):
# Evaluate output
critique = evaluator.critique(
task=task,
output=output,
criteria=[
"Correctness",
"Completeness",
"Clarity",
]
)
if critique["passes_all"]:
return output
# Refine based on critique
output = generator.refine(
task=task,
previous_output=output,
critique=critique["feedback"],
)
return output # Best effort after max iterations
"""
LangGraph Reflection
""" from langgraph.graph import StateGraph
def build_reflection_graph(): graph = StateGraph(ReflectionState)
# Nodes
graph.add_node("generate", generate_node)
graph.add_node("reflect", reflect_node)
graph.add_node("output", output_node)
# Edges
graph.add_edge("generate", "reflect")
graph.add_conditional_edges(
"reflect",
should_continue,
{
"continue": "generate", # Loop back
"end": "output",
}
)
return graph.compile()
def should_continue(state): if state["iteration"] >= 3: return "end" if state["score"] >= 0.9: return "end" return "continue" """
Separate Evaluator (More Robust)
"""
Use different model for evaluation to avoid self-bias
generator = ChatOpenAI(model="gpt-4o") evaluator = ChatOpenAI(model="gpt-4o-mini") # Different perspective
Or use specialized evaluators
from langchain.evaluation import load_evaluator evaluator = load_evaluator("criteria", criteria="correctness") """
Guardrailed Autonomy
Constrained agents with safety boundaries
When to use: Production systems, critical operations
GUARDRAILED AUTONOMY:
""" Production agents need multiple safety layers:
- Input validation
- Action constraints
- Output validation
- Cost limits
- Human escalation
- Rollback capability """
Multi-Layer Guardrails
""" class GuardedAgent: def init(self, agent, config): self.agent = agent self.max_cost = config.get("max_cost_usd", 1.0) self.max_steps = config.get("max_steps", 10) self.allowed_actions = config.get("allowed_actions", []) self.require_approval = config.get("require_approval", [])
async def execute(self, goal):
total_cost = 0
steps = 0
while steps < self.max_steps:
# Get next action
action = await self.agent.plan_next(goal)
# Validate action is allowed
if action.name not in self.allowed_actions:
raise ActionNotAllowedError(action.name)
# Check if approval needed
if action.name in self.require_approval:
approved = await self.request_human_approval(action)
if not approved:
return {"status": "rejected", "action": action}
# Estimate cost
estimated_cost = self.estimate_cost(action)
if total_cost + estimated_cost > self.max_cost:
raise CostLimitExceededError(total_cost)
# Execute with rollback capability
checkpoint = await self.save_checkpoint()
try:
result = await self.agent.execute(action)
total_cost += self.actual_cost(action)
steps += 1
except Exception as e:
await self.rollback_to(checkpoint)
raise
if result.is_complete:
break
return {"status": "complete", "total_cost": total_cost}
"""
Least Privilege Principle
"""
Define minimal permissions per task type
TASK_PERMISSIONS = { "research": ["web_search", "read_file"], "coding": ["read_file", "write_file", "run_tests"], "admin": ["all"], # Rarely grant this }
def create_scoped_agent(task_type): allowed = TASK_PERMISSIONS.get(task_type, []) tools = [t for t in ALL_TOOLS if t.name in allowed] return Agent(tools=tools) """
Cost Control
"""
Context length grows quadratically in cost
Double context = 4x cost
def trim_context(messages, max_tokens=4000): # Keep system message and recent messages system = messages[0] recent = messages[-10:]
# Summarize middle if needed
if len(messages) > 11:
middle = messages[1:-10]
summary = summarize(middle)
return [system, summary] + recent
return messages
"""
Durable Execution Pattern
Agents that survive failures and resume
When to use: Long-running tasks, production systems, multi-day processes
DURABLE EXECUTION:
""" Production agents must:
- Survive server restarts
- Resume from exact point of failure
- Handle hours/days of runtime
- Allow human intervention mid-process
LangGraph 1.0 provides this natively. """
LangGraph Checkpointing
""" from langgraph.checkpoint.postgres import PostgresSaver from langgraph.graph import StateGraph
Production checkpointer (not MemorySaver!)
checkpointer = PostgresSaver.from_conn_string( os.environ["POSTGRES_URL"] )
Build graph with checkpointing
graph = StateGraph(AgentState)
... add nodes and edges ...
agent = graph.compile(checkpointer=checkpointer)
Each invocation saves state
config = {"configurable": {"thread_id": "long-task-789"}}
Start task
agent.invoke({"goal": complex_goal}, config)
If server dies, resume later:
state = agent.get_state(config) if not state.is_complete: agent.invoke(None, config) # Continues from checkpoint """
Human-in-the-Loop Interrupts
"""
Pause at specific nodes
agent = graph.compile( checkpointer=checkpointer, interrupt_before=["critical_action"], # Pause before interrupt_after=["validation"], # Pause after )
First invocation pauses at interrupt
result = agent.invoke({"goal": goal}, config)
Human reviews state
state = agent.get_state(config) if human_approves(state): # Continue from pause point agent.invoke(None, config) else: # Modify state and continue agent.update_state(config, {"approved": False}) agent.invoke(None, config) """
Time-Travel Debugging
"""
LangGraph stores full history
history = list(agent.get_state_history(config))
Go back to any previous state
past_state = history[5] agent.update_state(config, past_state.values)
Replay from that point with modifications
agent.invoke(None, config) """
Sharp Edges
Error Probability Compounds Exponentially
Severity: CRITICAL
Situation: Building multi-step autonomous agents
Symptoms: Agent works in demos but fails in production. Simple tasks succeed, complex tasks fail mysteriously. Success rate drops dramatically as task complexity increases. Users lose trust.
Why this breaks: Each step has independent failure probability. A 95% success rate per step sounds great until you realize:
- 5 steps: 77% success (0.95^5)
- 10 steps: 60% success (0.95^10)
- 20 steps: 36% success (0.95^20)
This is the fundamental limit of autonomous agents. Every additional step multiplies failure probability.
Recommended fix:
Reduce step count
Combine steps where possible
Prefer fewer, more capable steps over many small ones
Increase per-step reliability
Use structured outputs (JSON schemas)
Add validation at each step
Use better models for critical steps
Design for failure
class RobustAgent: def execute_with_retry(self, step, max_retries=3): for attempt in range(max_retries): try: result = step.execute() if self.validate(result): return result except Exception as e: if attempt == max_retries - 1: raise self.log_retry(step, attempt, e)
Break into checkpointed segments
Human review at each segment
Resume from last good checkpoint
API Costs Explode with Context Growth
Severity: CRITICAL
Situation: Running agents with growing conversation context
Symptoms: $47 to close a single support ticket. Thousands in surprise API bills. Agents getting slower as they run longer. Token counts exceeding model limits.
Why this breaks: Transformer costs scale quadratically with context length. Double the context, quadruple the compute. A long-running agent that re-sends its full conversation each turn can burn money exponentially.
Most agents append to context without trimming. Context grows:
- Turn 1: 500 tokens → $0.01
- Turn 10: 5000 tokens → $0.10
- Turn 50: 25000 tokens → $0.50
- Turn 100: 50000 tokens → $1.00+ per message
Recommended fix:
Set hard cost limits
class CostLimitedAgent: MAX_COST_PER_TASK = 1.00 # USD
def __init__(self):
self.total_cost = 0
def before_call(self, estimated_tokens):
estimated_cost = self.estimate_cost(estimated_tokens)
if self.total_cost + estimated_cost > self.MAX_COST_PER_TASK:
raise CostLimitExceeded(
f"Would exceed ${self.MAX_COST_PER_TASK} limit"
)
def after_call(self, response):
self.total_cost += self.calculate_actual_cost(response)
Trim context aggressively
def trim_context(messages, max_tokens=4000): # Keep: system prompt + last N messages # Summarize: everything in between if count_tokens(messages) <= max_tokens: return messages
system = messages[0]
recent = messages[-5:]
middle = messages[1:-5]
if middle:
summary = summarize(middle) # Compress history
return [system, summary] + recent
return [system] + recent
Use streaming to track costs in real-time
Alert at 50% of budget, halt at 90%
Demo Works But Production Fails
Severity: CRITICAL
Situation: Moving from prototype to production
Symptoms: Impressive demo to stakeholders. Months of failure in production. Works for the founder's use case, fails for real users. Edge cases overwhelm the system.
Why this breaks: Demos show the happy path with curated inputs. Production means:
- Unexpected inputs (typos, ambiguity, adversarial)
- Scale (1000 users, not 3)
- Reliability (99.9% uptime, not "usually works")
- Edge cases (the 1% that breaks everything)
The methodology is questionable, but the core problem is real. The gap between a working demo and a reliable production system is where projects die.
Recommended fix:
Test at scale before production
Run 1000+ test cases, not 10
Measure P95/P99 success rate, not average
Include adversarial inputs
Build observability first
import structlog logger = structlog.get_logger()
class ObservableAgent: def execute(self, task): with logger.bind(task_id=task.id): logger.info("task_started") try: result = self._execute(task) logger.info("task_completed", result=result) return result except Exception as e: logger.error("task_failed", error=str(e)) raise
Have escape hatches
Human takeover when confidence < threshold
Graceful degradation to simpler behavior
"I don't know" is a valid response
Deploy incrementally
1% of traffic, then 10%, then 50%
Monitor error rates at each stage
Agent Fabricates Data When Stuck
Severity: HIGH
Situation: Agent can't complete task with available information
Symptoms: Agent invents plausible-looking data. Fake restaurant names on expense reports. Made-up statistics in reports. Confident answers that are completely wrong.
Why this breaks: LLMs are trained to be helpful and produce plausible outputs. When stuck, they don't say "I can't do this" - they fabricate. Autonomous agents compound this by acting on fabricated data without human review.
The agent that fabricated expense entries was trying to meet its goal (complete the expense report). It "solved" the problem by inventing data.
Recommended fix:
Validate against ground truth
def validate_expense(expense): # Cross-check with external sources if expense.restaurant: if not verify_restaurant_exists(expense.restaurant): raise ValidationError("Restaurant not found")
# Check for suspicious patterns
if expense.amount == round(expense.amount, -1):
flag_for_review("Suspiciously round amount")
Require evidence
system_prompt = ''' For every factual claim, cite the specific tool output that supports it. If you cannot find supporting evidence, say "I could not verify this" rather than guessing. '''
Use structured outputs
from pydantic import BaseModel
class VerifiedClaim(BaseModel): claim: str source: str # Must reference tool output confidence: float
Detect uncertainty
Train to output confidence scores
Flag low-confidence outputs for human review
Never auto-execute on uncertain data
Integration Is Where Agents Die
Severity: HIGH
Situation: Connecting agent to external systems
Symptoms: Works with mock APIs, fails with real ones. Rate limits cause crashes. Auth tokens expire mid-task. Data format mismatches. Partial failures leave systems in inconsistent state.
Why this breaks: The companies promising "autonomous agents that integrate with your entire tech stack" haven't built production systems at scale. Real integrations have:
- Rate limits (429 errors mid-task)
- Auth complexity (OAuth refresh, token expiry)
- Data format variations (API v1 vs v2)
- Partial failures (webhook received, processing failed)
- Eventual consistency (data not immediately available)
Recommended fix:
Build robust API clients
from tenacity import retry, stop_after_attempt, wait_exponential
class RobustAPIClient: @retry( stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=4, max=60) ) async def call(self, endpoint, data): response = await self.client.post(endpoint, json=data) if response.status_code == 429: retry_after = response.headers.get("Retry-After", 60) await asyncio.sleep(int(retry_after)) raise RateLimitError() return response
Handle auth lifecycle
class TokenManager: def init(self): self.token = None self.expires_at = None
async def get_token(self):
if self.is_expired():
self.token = await self.refresh_token()
return self.token
def is_expired(self):
buffer = timedelta(minutes=5) # Refresh early
return datetime.now() > (self.expires_at - buffer)
Use idempotency keys
Every external action should be idempotent
If agent retries, external system handles duplicate
Design for partial failure
Each step is independently recoverable
Checkpoint before external calls
Rollback capability for each integration
Agent Takes Dangerous Actions
Severity: HIGH
Situation: Agent with broad permissions
Symptoms: Agent deletes production data. Sends emails to wrong recipients. Makes purchases without approval. Modifies settings it shouldn't. Actions that can't be undone.
Why this breaks: Agents optimize for their goal. Without guardrails, they'll take the shortest path - even if that path is destructive. An agent told to "clean up the database" might interpret that as "delete everything."
Broad permissions + autonomy + goal optimization = danger.
Recommended fix:
Least privilege principle
PERMISSIONS = { "research_agent": ["read_web", "read_docs"], "code_agent": ["read_file", "write_file", "run_tests"], "email_agent": ["read_email", "draft_email"], # NOT send "admin_agent": ["all"], # Rarely used }
Separate read/write permissions
Agent can read anything
Write requires explicit approval
Dangerous actions require confirmation
DANGEROUS_ACTIONS = [ "delete_*", "send_email", "transfer_money", "modify_production", "revoke_access", ]
async def execute_action(action): if matches_dangerous_pattern(action): approval = await request_human_approval(action) if not approval: return ActionRejected(action) return await actually_execute(action)
Dry-run mode for testing
Agent describes what it would do
Human approves the plan
Then agent executes
Audit logging for everything
Every action logged with context
Who authorized it
What changed
How to reverse it
Agent Runs Out of Context Window
Severity: MEDIUM
Situation: Long-running agent tasks
Symptoms: Agent forgets earlier instructions. Contradicts itself. Loses track of the goal. Starts repeating itself. Model errors about token limits.
Why this breaks: Every message, observation, and thought consumes context. Long tasks exhaust the window. When context is truncated:
- System prompt gets dropped
- Early important context lost
- Agent loses coherence
Recommended fix:
Track context usage
class ContextManager: def init(self, max_tokens=100000): self.max_tokens = max_tokens self.messages = []
def add(self, message):
self.messages.append(message)
self.maybe_compact()
def maybe_compact(self):
if self.token_count() > self.max_tokens * 0.8:
self.compact()
def compact(self):
# Always keep: system prompt
system = self.messages[0]
# Always keep: last N messages
recent = self.messages[-10:]
# Summarize: everything else
middle = self.messages[1:-10]
if middle:
summary = summarize_messages(middle)
self.messages = [system, summary] + recent
Use external memory
Don't keep everything in context
Store in vector DB, retrieve when needed
See agent-memory-systems skill
Hierarchical summarization
Recent: full detail
Medium: key points
Old: compressed summary
Can't Debug What You Can't See
Severity: MEDIUM
Situation: Agent fails mysteriously
Symptoms: "It just didn't work." No idea why agent failed. Can't reproduce issues. Users report problems you can't explain. Debugging is guesswork.
Why this breaks: Agents make dozens of internal decisions. Without visibility into each step, you're blind to failure modes. Production debugging without traces is impossible.
Recommended fix:
Structured logging
import structlog
logger = structlog.get_logger()
class TracedAgent: def think(self, context): with logger.bind(step="think"): thought = self.llm.generate(context) logger.info("thought_generated", thought=thought, tokens=count_tokens(thought) ) return thought
def act(self, action):
with logger.bind(step="act", action=action.name):
logger.info("action_started")
try:
result = action.execute()
logger.info("action_completed", result=result)
return result
except Exception as e:
logger.error("action_failed", error=str(e))
raise
Use LangSmith or similar
from langsmith import trace
@trace def agent_step(state): # Automatically traced with inputs/outputs return next_state
Save full traces
Every step, every decision
Inputs and outputs
Latency at each step
Token usage
Validation Checks
Agent Loop Without Step Limit
Severity: ERROR
Autonomous agents must have maximum step limits
Message: Agent loop without step limit. Add max_steps to prevent infinite loops.
No Cost Tracking or Limits
Severity: ERROR
Agents should track and limit API costs
Message: Agent uses LLM without cost tracking. Add cost limits to prevent runaway spending.
Agent Without Timeout
Severity: WARNING
Long-running agents need timeouts
Message: Agent invocation without timeout. Add timeout to prevent hung tasks.
MemorySaver Used in Production
Severity: ERROR
MemorySaver is for development only
Message: MemorySaver is not persistent. Use PostgresSaver or SqliteSaver for production.
Long-Running Agent Without Checkpointing
Severity: WARNING
Agents that run multiple steps need checkpointing
Message: Multi-step agent without checkpointing. Add checkpointer for durability.
Agent Without Thread ID
Severity: WARNING
Checkpointed agents need unique thread IDs
Message: Agent invocation without thread_id. State won't persist correctly.
Using Agent Output Without Validation
Severity: WARNING
Agent outputs should be validated before use
Message: Agent output used without validation. Validate before acting on results.
Agent Without Structured Output
Severity: INFO
Structured outputs are more reliable
Message: Consider using structured outputs (Pydantic) for more reliable parsing.
Agent Without Error Recovery
Severity: WARNING
Agents should handle and recover from errors
Message: Agent call without error handling. Add try/catch or error handler.
Destructive Actions Without Rollback
Severity: WARNING
Actions that modify state should be reversible
Message: Destructive action without rollback capability. Save state before modification.
Collaboration
Delegation Triggers
- user needs multi-agent coordination -> multi-agent-orchestration (Multiple agents working together)
- user needs to test/evaluate agent -> agent-evaluation (Benchmarking and testing)
- user needs tools for agent -> agent-tool-builder (Tool design and implementation)
- user needs persistent memory -> agent-memory-systems (Long-term memory architecture)
- user needs workflow automation -> workflow-automation (When agent is overkill for the task)
- user needs computer control -> computer-use-agents (GUI automation, screen interaction)
Related Skills
Works well with: agent-tool-builder, agent-memory-systems, multi-agent-orchestration, agent-evaluation
When to Use
- User mentions or implies: autonomous agent
- User mentions or implies: autogpt
- User mentions or implies: babyagi
- User mentions or implies: self-prompting
- User mentions or implies: goal decomposition
- User mentions or implies: react pattern
- User mentions or implies: agent loop
- User mentions or implies: self-correcting agent
- User mentions or implies: reflection agent
- User mentions or implies: langgraph
- User mentions or implies: agentic ai
- User mentions or implies: agent planning
Limitations
- Use this skill only when the task clearly matches the scope described above.
- Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
- Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.