Open Specification

Agent Spending Policy Specification

A vendor-neutral format for AI agent spending control. From single-agent budgets to multi-agent mission coordination.

ASPS v1

Stable

Single-agent spending policies. Budgets, limits, categories, schedules, auto-approval. Implemented in LetAgentPay.

  • 9 deterministic checks
  • Account-level budget rules
  • Fund holding for pending requests
  • JSON Schema validation
Read v1 spec

ASPS v2

Draft

Multi-agent mission coordination. Shared budgets, phases, dependencies, dynamic allocation, risk scoring.

  • Mission-level budget orchestration
  • 5 allocation strategies (incl. competitive)
  • Phase transitions and constraints
  • Anomaly detection via risk scoring

v1: Single-agent policy

Stable
{
  "version": "1.0",
  "daily_limit": 500,
  "weekly_limit": 2000,
  "per_request_limit": 200,
  "allowed_categories": ["groceries", "food_delivery", "subscriptions"],
  "schedule": {
    "timezone": "America/New_York",
    "default": { "allow": "08:00-22:00" },
    "overrides": [
      { "days": ["sat", "sun"], "allow": "10:00-18:00", "daily_limit": 100 },
      { "days": ["wed"], "deny": true }
    ]
  },
  "auto_approve": {
    "enabled": true,
    "max_amount": 50,
    "categories": ["groceries", "food_delivery"]
  }
}

This policy says:

  • Up to $500/day, $2,000/week, $200 per purchase
  • Only groceries, food delivery, and subscriptions
  • Weekdays 08:00–22:00, weekends 10:00–18:00, no Wednesdays
  • Weekend daily limit reduced to $100
  • Groceries and food delivery under $50 — auto-approved

9 standard checks

Every spending request is evaluated against these checks, in order. All checks run (no short-circuit) to provide a complete report.

1Agent status

Is the agent active?

2Category

Is this spending category allowed?

3Per-request limit

Does this single purchase exceed the cap?

4Schedule

Is spending allowed right now?

5Daily limit

Would this push today's total over the daily cap?

6Weekly limit

Same check for the current week

7Monthly limit

Same check for the current month

8Total budget

Would this exceed the agent's overall budget?

9Account rules

Cross-agent budget rules at the account level

v2: Mission coordination

Draft

Multiple agents, one goal, shared budget. Phases unlock as work progresses. Budget flows dynamically based on actual spending.

{
  "version": "2.0",
  "name": "Barcelona trip, May 15-22",
  "budget": 5000,
  "currency": "USD",
  "agents": {
    "flights": { "policy": { "allowed_categories": ["flights"] } },
    "hotel":   { "policy": { "allowed_categories": ["accommodation"] } },
    "fun":     { "policy": { "allowed_categories": ["entertainment"] } }
  },
  "phases": [
    { "name": "booking", "agents": ["flights", "hotel"],
      "allocation": { "type": "share", "percent": 70 } },
    { "name": "activities", "agents": ["fun"],
      "allocation": { "type": "remaining" } }
  ],
  "constraints": [
    { "type": "dependency", "agent": "hotel",
      "requires": "flights", "condition": "confirmed" }
  ]
}

This mission says:

  • $5,000 total budget for a Barcelona trip
  • 3 agents: flights, hotel, entertainment
  • Booking phase gets 70% ($3,500), dynamic reallocation
  • Hotel can’t book until flights are confirmed
  • Activities phase gets whatever remains
  • Cheap flights = more budget for hotel and fun

What v2 adds

Mission coordination

Multiple agents work toward a common goal with a shared budget. Phases, dependencies, and dynamic allocation.

5 allocation strategies

Fixed, percentage share, remaining, per-agent, and competitive. Budget flows dynamically based on actual spending.

Phase transitions

Gates, escalation, timeouts. Agents unlock as phases complete. Rollback on failure.

Cross-agent constraints

Dependencies, combined limits, conditional rules, exclusions, priority ordering between agents.

Risk scoring

Anomaly detection based on agent behavior history. Flag unusual amounts, categories, timing, or frequency.

Escalation model

Same agent, increasing authority. L1 can offer 5% discount, L2 can refund $100, L3 can compensate $500.

Architecture

ASPS v2 (missions, phases, allocation, risk scoring)
ASPS v1 (per-agent policy: limits, categories, schedule)
9 standard checks → pass/fail

v2 is a layer above v1, not a replacement. Every agent in a mission still uses a v1 policy. Agents can operate independently (v1 only) or as part of a mission (v2).

We don't move money. We decide if it should move.

ASPS is an open specification. Use it with any payment provider and any AI framework.