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% b
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
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
'''
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.
'''
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
"""
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
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:]
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)
"""
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 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:
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."
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.
