Where the time goes now, why review does not catch it, and what to try next.
FROM THE AI-READY INFRASTRUCTURE PLAYBOOK10 MIN
The problem
Output is up, but nothing reaches customers faster.
Writing code was one stage of the work, and it is the one that sped up. Everything after it still runs at the old speed, and each request still needs a stack of infrastructure to run.
93%
of organisations have had at least one AI-caused infrastructure incident
45%
of deployments containing AI-generated code lead to problems
4.6x
longer wait for a first review on AI-authored pull requests
Every check passed. The pager went off 16 hours later.
You only find out in production, so an agent gets one attempt.
The fix
Give infrastructure the same loop, by deriving it from the code.
Encore is one platform built this way. Whoever makes the change can run it and see the result before production.
Your code & agents
Typed declarations in TypeScript or Go
Encore
Provisions the infrastructure
Your cloud
Runs in your own AWS or GCP
Infrastructure from application code
What that looks like in practice
Resources are declared as typed objects in ordinary TypeScript or Go, and the platform derives provisioning, permissions, environments and tracing from them, with sensible defaults you can override in the dashboard.
const users = new SQLDatabase("users", { migrations: "./migrations" });
const signups = new Topic<SignupEvent>("signups", { deliveryGuarantee: "at-least-once" });
Local
On your machine
A database and queue running locally, not mocked.
Preview
One per pull request
Reviewers open the running change, not a plan.
Production
Your own AWS or GCP
Same declarations, tracing already attached.
All three come from the same declarations, so a change runs and gets fixed before anyone reviews it.
Where we stand
How late do we find out today?
Four possible answers. Whichever one is true for us is the size of the problem, because the later we find out, the more AI-generated volume makes it hurt.
At build timeRare. The loop is already closed.
In code reviewA person reading intent. The expensive mistakes get through.
In staging, sometimesShared, slow to reset, and skipped when we are in a hurry.
In productionGets worse the better AI adoption goes.
The proposal
Build one new service this way, and compare the numbers
Pick a service we were going to build anyway and build it on a platform that derives infrastructure from application code, Encore for example. Everything we run today stays where it is, and nothing gets migrated. A sprint to build, then it runs on its own, and we decide on our own numbers.
A sprintBuild the service
ThenLeave it running in production
AfterCompare
The four numbers, taken before we start and again once it has run
Lead time
Commit to running in production
Review wait
Opened to first review
Change failure rate
Deploys needing a fix
Mistakes caught late
How many reached staging or production
Sources, benchmarks and the full readiness diagnostic are in the report.Read the full report