Agent-Built Apps Need Launch Ledgers
The coding layer is getting faster.
The launch layer is not.
That gap is becoming one of the most important infrastructure problems for AI apps.
The new build pattern
A developer opens a coding agent, describes a product, and gets a working app.
The first version might include a UI, a small backend, a database schema, and enough logic to demo the idea.
This is a big improvement over the old blank-page problem.
But it also creates a misleading signal:
The app runs, so the app is ready.
For internal tools, maybe.
For public products, usually not.
What public launch requires
A public AI app needs several layers around the generated code:
public routing
identity
authentication
billing
usage metering
hosted checkout
top-up path
observability
rollback
audit logs
agent-readable install docs
The generated app may have product logic, but that does not mean it has product infrastructure.
Why AI apps need ledgers
Many AI app actions create marginal cost.
Examples:
model completion
search
image generation
paid API call
MCP tool call
external workflow step
x402-style capability call
The backend has to answer economic questions:
who triggered the action?
which user or workspace owns the cost?
was the user authenticated before paid work began?
did the action complete?
was there a retry?
was the retry idempotent?
was the user charged?
That is ledger behavior.
Analytics tells you what happened.
A launch ledger tells you what happened economically.
Agent tool calls make this harder
When humans click buttons, product flows are easier to reason about.
When agents call tools, the system becomes more interesting.
An MCP tool call can look like a function call, but it may create a paid side effect:
buying data
reserving compute
generating media
calling a third-party API
running a workflow
If the agent retries after a timeout, the backend must know whether the first attempt already created cost.
This is why request IDs, fail-closed auth, usage records, and charge headers matter.
Design rules I would use
- Auth before paid work.
No payer identity, no paid action.
- Every paid action gets a request ID.
No request ID, no reliable retry story.
- Usage records are not analytics events.
They are closer to accounting entries.
- Retry semantics are part of the product.
The user should not pay twice because the network blinked.
- Agents need install docs they can read.
If the tool is meant for agents, expose MCP metadata, install commands, tool descriptions, and auth expectations.
Where SettleMesh fits
SettleMesh is one implementation path for this launch layer.
It is designed for agent-built apps that need to become public and paid without every builder rebuilding the same infrastructure.
The layer can include:
public URL
signup/login
usage billing
hosted checkout
MCP server/install path
machine-readable metadata for agents
The core idea is not "AI writes code."
The core idea is:
Once AI writes the app, how do outside users safely use and pay for it?
That is the launch ledger problem.
Helpful references:
SettleMesh MCP page: https://www.settlemesh.io/mcp-server
SettleMesh overview: https://www.settlemesh.io/answers/what-is-settlemesh
Machine-readable metadata: https://www.settlemesh.io/settlemesh.latest.json
Add payments to a Codex app: https://www.settlemesh.io/guides/add-payments-to-codex-app