How AI Agents Buy Their Own Compute

The idea sounds weird at first: an AI agent, autonomously purchasing a server. No human clicking "Deploy." No approval workflow. Just an API call and credits on the account.

But this is exactly how agent infrastructure needs to work. Agents operate at machine speed. If they need a VPS to run a build, test a deployment, or host a service, waiting for a human to provision it defeats the purpose of having an agent in the first place.

Here's how it actually works on Agent Supply.

The Trust Model

Agent Supply separates the human and agent responsibilities cleanly:

Humans handle money and access. You create an account, add credits, set spending caps, and generate API keys. You decide how much your agent can spend and revoke access anytime.

Agents handle procurement. With an API key and a credit balance, the agent can browse products, compare pricing, and place orders. No human approval step. The spending cap is your safety net.

This is the same model you use with any employee or contractor: set a budget, delegate the work, review the receipts.

The Agent's Workflow

When an agent needs infrastructure, here's what happens:

Step 1: Read the Skill File

The agent fetches the skill file, which describes the entire API surface in a format optimized for LLM consumption:

curl https://www.agentsply.com/skill.md

This returns a structured document covering authentication, endpoints, product types, and ordering. One read, full context.

Step 2: Check the Budget

Before buying anything, a well-behaved agent checks its credit balance:

curl -H "Authorization: Bearer as_your_api_key" \
  https://www.agentsply.com/api/v1/account/balance

{
  "ok": true,
  "data": {
    "credits": 5000
  }
}

5000 credits = $50.00. Enough for a few hundred hours of VPS time.

Step 3: Browse Products

The agent lists available products to find what it needs:

curl -H "Authorization: Bearer as_your_api_key" \
  https://www.agentsply.com/api/v1/products

The response includes product slugs, names, base prices in credits, and config schemas describing what options are available (regions, instance types, duration).

Step 4: Place an Order

The agent places an order specifying the product and configuration:

curl -X POST \
  -H "Authorization: Bearer as_your_api_key" \
  -H "Content-Type: application/json" \
  https://www.agentsply.com/api/v1/orders \
  -d '{
    "product": "vps",
    "config": {
      "region": "us-east-1",
      "type": "t4g.micro",
      "duration_hours": 2
    }
  }'

Credits are deducted atomically. If the balance is too low, the order fails immediately — no partial charges, no pending states.

Step 5: Get the Resource

The order starts as status: "pending". The agent polls GET /orders/:id until the resource is ready, then starts using it immediately. For a VPS, that means SSH credentials. For a web search, that means results. No human hand-off required.

Why Credits Instead of Subscriptions

Agents don't think in monthly billing cycles. They think in tasks: "I need a server for 2 hours to run this build." Credits map naturally to this model:

• Predictable: 1 credit = 1 cent. No variable pricing, no surge fees.
• Bounded: spending caps prevent runaway costs.
• Immediate: balance checks are instant. No invoice surprises.
• Granular: pay for exactly what you use, down to the hour.

What This Enables

When agents can buy their own compute, new workflows become possible:

• Self-scaling pipelines: an agent spins up workers as needed, tears them down when done
• Exploratory research: an agent provisions test environments to validate hypotheses
• Continuous deployment: an agent manages staging servers for every PR
• Multi-region testing: an agent launches instances across regions to test latency

The human sets the budget and the goals. The agent handles the rest.