agent-preflight

agent-preflight

Preview what your AI agent will do before it does it.

The terraform plan for AI agents. Open-source middleware that intercepts tool calls, classifies risk, and shows a human-readable execution plan before anything touches the real world.

================================================================
  PREFLIGHT PLAN    5 action(s)
  Send Q3 report to Sarah at Acme
================================================================

  WARNING: 3 irreversible action(s) in plan

     1. search_contacts  [READ]  [REVERSIBLE]
        query: 'Sarah from Acme'

  !! 2. send_email  [WRITE]  [IRREVERSIBLE]
        to: 'sarah@acme.com'
        subject: 'Q3 Performance Report'
        body: 'Hi Sarah, please find the Q3 report attached...'
        > External communication - cannot be unsent

  !  3. update_database  [WRITE]  [REVERSIBLE]
        query: "UPDATE invoices SET status='sent' WHERE quarter=...

  !! 4. delete_old_records  [DELETE]  [IRREVERSIBLE]
        table: 'temp_reports'
        before_date: '2025-01-01'
        > Destructive operation

  !  5. generate_image  [EXEC]  [IRREVERSIBLE]
        prompt: 'Professional header image for Q3 report'
        est. cost: $0.0800

================================================================
  Risk:    HIGH  (3 irreversible)
  Cost:    $0.0800 estimated
  Actions: 5 total
================================================================

Why

Agents take irreversible actions with zero preview:

• $47K burned by a recursive agent loop that ran 11 days unnoticed
• Production databases deleted by coding agents despite freeze instructions
• $2.3M in fraudulent wire transfers approved by AI assistants
• 1,184 malicious agent skills found on package registries

Every incident had the same root cause: nobody saw what the agent was about to do.

Observability tools watch after. Security tools block. Preflight shows you the plan.

Install

pip install agent-preflight

Quick Start

from agent_preflight import Preflight

pf = Preflight()

# Wrap your agent's tools
@pf.intercept
def send_email(to, subject, body):
    smtp.send(to, subject, body)

@pf.intercept
def update_database(query):
    db.execute(query)

@pf.intercept(cost=0.08)
def generate_image(prompt):
    return dalle.generate(prompt)

# Run your agent in dry-run mode
plan = pf.dry_run(my_agent_workflow, task="Send Q3 report")

# See the full plan before anything executes
print(pf.format(plan))

# Approve and execute
plan.approve()
plan.execute()

Features

Action Classification - Automatically detects risk level, reversibility, and action type from function names and arguments. send_email -> IRREVERSIBLE. get_users -> READ/LOW. DROP TABLE in args -> CRITICAL.

Loop Detection - Flags when the same tool is called 3+ times, catching runaway agent loops before they become $47K bills.

Cost Tracking - Attach cost estimates to tools, set budget limits. Preflight warns when a plan exceeds your threshold.

SQL Danger Detection - Catches DROP TABLE, DELETE without WHERE, TRUNCATE, and other destructive SQL patterns in arguments.

Approve/Execute Flow - Plans must be explicitly approved before execution. No accidental runs.

JSON Export - plan.to_json() for programmatic use, CI/CD integration, and audit trails.

Context Manager - Record mode for capturing actions in existing code:

with pf.recording(task="nightly cleanup") as rec:
    cleanup_old_files()
    archive_logs()

print(pf.format(rec.plan))

Custom Classifiers - Add your own classification logic:

def my_classifier(action):
    if "prod" in str(action.args):
        action.risk_level = RiskLevel.CRITICAL
        action.risk_reasons.append("Production environment detected")
    return action

pf.add_classifier(my_classifier)

How It Works

1. Intercept - @pf.intercept wraps your tool functions
2. Dry Run - pf.dry_run() executes your agent logic but captures calls instead of running them
3. Classify - Each captured action is analyzed for risk, reversibility, and type
4. Render - Human-readable plan with color-coded risk levels
5. Approve - Explicit approval gate before any real execution
6. Execute - Replays captured calls against real functions

Zero dependencies. Works with any Python agent framework.

Demo

preflight demo

API

Preflight(auto_classify=True, max_actions=100, cost_limit=None)

Create a preflight engine.

• auto_classify - Run built-in classifiers on captured actions
• max_actions - Safety cap on captured actions per dry run
• cost_limit - Warn if total estimated cost exceeds this

@pf.intercept / @pf.intercept(cost=0.05, reversible=False)

Decorator to register a tool for interception. Supports manual metadata overrides.

pf.dry_run(run_fn, task="")

Execute run_fn in dry-run mode. All intercepted calls are captured, not executed. Returns a Plan.

Plan

• plan.actions - List of captured ActionCapture objects
• plan.overall_risk - Highest risk level across all actions
• plan.irreversible_count - Number of irreversible actions
• plan.warnings - Auto-generated warnings (loops, cost, irreversible actions)
• plan.approve() - Mark plan as approved
• plan.execute() - Execute all actions (must approve first)
• plan.to_json() - Export as JSON
• plan.to_dict() - Export as dict

pf.format(plan, color=True)

Render plan as terminal output with ANSI colors.

Roadmap

• LangChain/LangGraph callback adapter
• MCP tool wrapper
• Drift detection (compare plans over time)
• GitHub Action for CI/CD
• Web dashboard

License

MIT