Before any payment can happen, something must decide that a payment should happen. That's Layer 1 — and it's almost entirely unsolved. No standard exists for how agents budget, prioritize, and authorize their own spending.
Layer 1 is the "brain" of agent commerce. It's where an LLM reasons about whether a purchase is worth making, what the budget ceiling is, which of several options to choose, and when to escalate to a human. Today this logic lives inside individual agent systems with no shared primitives or standards.
The unanswered questions
Q1
How does an agent know how much it's allowed to spend?
Budget context is currently injected via system prompt. No structured format exists for machine-readable spend budgets.
Q2
How does an agent choose between options at different price points?
Price-quality tradeoffs require preference modeling. Most agents default to the cheapest or the first option — not the best one.
Q3
When should the agent stop and ask for approval?
Escalation logic — when a purchase exceeds a threshold, is unusual, or is irreversible — is ad hoc in every agent today.
Q4
How does an agent learn from past spending?
Memory of past purchases, vendor reliability, and price trends should inform future decisions. No agent memory standard exists for financial context.
Q5
Who is liable when the agent makes a bad purchase decision?
If an agent overbids on a compute job or buys from a fraudulent vendor, who holds accountability — the agent, the user, the framework?
Q6
How do multi-agent systems coordinate spending?
When an orchestrator spawns sub-agents, how are budgets allocated and tracked across the hierarchy? No protocol exists for agent treasury delegation.
Who is building at this layer
OpenAI's agent framework gives GPT-4o and o3 the ability to take multi-step actions including payments via tool calling. Spend intent emerges from the model's reasoning — no explicit budget primitive. The model decides when to call a payment tool based on context in the system prompt.
OpenAI Agents docs →Claude's tool use system allows payment tools (Stripe, x402, etc.) to be called when the model decides a purchase is warranted. Claude's extended thinking capability allows more deliberate cost-benefit reasoning before triggering a payment. Budget context injected via system prompt.
Claude tool use docs →Built agents that take real-world actions in software interfaces — including navigating checkout flows and completing purchases. Adept's agents can handle the full purchase journey but the "should I buy this?" logic is task-specific rather than a generalized intent standard.
adept.ai →AutoGPT and similar frameworks (BabyAGI, AgentGPT) pioneered autonomous goal-directed agents. They generate purchase decisions from goal decomposition, but budget management is a solved-in-isolation problem — no shared standard across frameworks.
AutoGPT on GitHub →❗ The Biggest Opportunity in the Stack
Layer 1 is the only layer with no incumbents. Every other layer has Visa, Stripe, or a blockchain protocol staking a position. Layer 1 — the "should the agent spend?" logic — is a blank page. The company that defines a standard for agent spend intent (structured budget primitives, preference models, escalation protocols, multi-agent treasury delegation) creates the foundation all other layers depend on.
Think of it as the missing "financial operating system" for AI agents. Today every agent developer solves this in their system prompt. The abstraction that makes spend intent portable, auditable, and composable doesn't exist yet.