Feature Proposal: Milestone-Based Billing for AI Coding Agents
I would like to propose an alternative billing model for Google Antigravity and other AI coding agents:
Milestone-Based AI Billing — Pay for Accepted Progress, Not Failed Retries.
The Problem
With usage-based billing, customers pay for tokens or compute regardless of whether the AI agent succeeds.
A real coding workflow can easily become:
Build → Error → Fix → New Error → Retry → Fix Again → Test → Retry
The customer’s credits continue to decrease throughout this process.
This creates an uncomfortable situation: the customer is effectively paying for the AI agent to correct mistakes created by the agent itself.
For small prompts this may be acceptable, but for autonomous coding agents working on real projects, I believe a different billing option could provide much greater customer confidence.
My Proposal
Divide autonomous development work into clearly defined milestones.
For example, a customer deposits $100.
Before starting a milestone, Antigravity estimates its cost:
Available balance: $100
Milestone cost: $12
Instead of immediately consuming the $12, the system reserves it:
$88 Available + $12 Reserved
The agent then:
Defines Scope → Builds → Tests → Presents the Result
The customer reviews the milestone.
If Accepted
The reserved $12 becomes a finalized charge, and the agent proceeds to the next milestone.
If the Implementation Is Incorrect
The agent receives a limited Correction Allowance to fix its implementation without charging the customer the full milestone cost again.
This allowance could be based on:
-
A percentage of milestone compute
-
A limited number of correction attempts
-
A token allowance
-
Or a combination of these
Correction vs. New Request
There should be an important distinction:
AI implementation error: covered by the Correction Allowance.
Customer changes requirements or requests additional functionality: billed as new work.
An independent AI evaluator could compare the original specification with the delivered result to help determine the difference.
Why I Think This Matters
This model creates better incentives for both sides.
The customer has a predictable project budget.
Google still has secured funds through reserved credits.
Most importantly:
The better the AI performs on the first attempt, the more profitable the service becomes.
Instead of AI mistakes generating additional revenue through additional token consumption, better AI performance itself becomes economically valuable.
This could become especially important as AI products evolve from chat assistants into autonomous software engineering agents.
Suggested Workflow
Deposit → Define Milestone → Estimate → Reserve Credits → Build → Test → Review → Correct if Needed → Accept → Charge → Continue
Traditional token-based billing does not need to disappear.
This could simply be an optional billing mode for autonomous project work.
The core idea is:
Your credits should pay for accepted progress, not endless retries.
I believe a system like this could significantly increase trust in paid autonomous coding agents and make customers much more comfortable allocating larger budgets to AI-driven development.