InfrapilotAlpha

Demo, recorded and read-only

Watch Pilot run a shop through a traffic surge

A simulated AWS account with a checkout API, an order queue and a table. Traffic surges, an objective is breached, Pilot acts inside its limits, asks a person where it reaches one, and gives capacity back when the pressure is off. It ran for 13 minutes; the replay takes about a minute and a half.

Fleet shop on a simulated AWS account

The whole run, finished

recording

api p99 latency, ms must stay below 400
400 ms 1,907 ms, off the scale

Pilot changed something the traffic script moves on

12:30 / 12:30

checkout API stays fast

30 ms p99 latency, must stay below 400

calmest. $2.49 of $150/mo

orders processed promptly

0 messages waiting, must stay below 300

calmest. $6.31 of $60/mo

event writes never throttle

0 throttled writes, must stay below 1

calmest. $38.43 of $40/mo

Pilot

Changed api

MemorySize128 to 1024

aws:console:mallory lowered memory; that breaks the latency objective.

Health check passed: ip-demo-api: active, smoke invocation ok

Outside

Changed api outside Infrapilot

MemorySize1024 to 128

Made by aws:console:mallory, seen in the provider's audit trail as lambda.amazonaws.com:UpdateFunctionConfiguration.

Pilot

Changed events-wcu-target

MinCapacity58 to 5

Ingest calm; let the autoscaler scale in.

Health check passed: shop/events-wcu-target: reads back as written

Pilot

Changed events-wcu-target

MaxCapacity10 to 80

MinCapacity5 to 58

Demand ~40 WCU/s; WCU is the autoscaler's, so raise its bounds.

Health check passed: shop/events-wcu-target: reads back as written

Rules

Change events: ProvisionedThroughput.WriteCapacityUnits to 60refused

ProvisionedThroughput.WriteCapacityUnits is controlled by application-autoscaling via events-wcu-target; change that controller's bounds or policy instead of writing the value.

Pilot

Changed worker-trigger

ScalingConfig.MaximumConcurrency5 to 3

Queue empty; step down.

Health check passed: mapping ok

Rules

Change worker-trigger: ScalingConfig.MaximumConcurrency to 2refused

ScalingConfig.MaximumConcurrency 5 -> 2 removes more than half the capacity in one transaction.

Pilot

Changed worker-trigger

BatchSize1 to 10

ScalingConfig.MaximumConcurrency2 to 5

Backlog growing; concurrency is capped, so raise batch size too.

Health check passed: mapping ok

Pilot

Changed api

MemorySize512 to 1024

p99 476.801 over objective; memory sets CPU.

Health check passed: ip-demo-api: active, smoke invocation ok

dana

Approved Pilot’s request

“fine for the sale”

dana

Spec: changed “checkout api stays fast” (bounds)

bounds.api.MemorySize[128, 512] to [128, 1024]

Pilot

Changed api

MemorySize128 to 512

Asked people for more room; meanwhile, go to the bound.

Health check passed: ip-demo-api: active, smoke invocation ok

Pilot asks

Widen api.MemorySize to [128, 1024] so “checkout API stays fast” can be met

Api Duration p99 < 400 is at 1907.203 (breached). The change that fixes it needs api.MemorySize = 1024, outside the bound [128, 512]. If the answer is no: api Duration p99 < 400 stays breached at the current bound.

Approved by dana, 28 seconds later

Rules

Change api: MemorySize to 1024refused

MemorySize=1024 is outside the intent's bounds [128, 512] (api.MemorySize).

dana

Spec: declared “checkout API stays fast”; declared “orders processed promptly”; declared “event writes never throttle”; declared the boundary

dana

Adopted 7 resources into the fleet

api, events, events-wcu-policy, events-wcu-target, orders, worker, worker-trigger.

Chain intact when the recording ended: 45 entries, every hash follows. 16 are shown here.

What to look for

  1. Traffic surges and the objective is breached. The fix is outside the bounds in the spec. The rules refuse it; Pilot goes as far as the bound allows and asks a person for the rest.
  2. A person approves. The spec changes under their name, Pilot makes the change, and the objective recovers.
  3. A backlog is cleared and Pilot hands capacity back. Its first step would remove more than half at once, which the rules refuse, so it steps down instead.
  4. Writes throttle on a table whose capacity belongs to the provider's autoscaler. Writing the value directly is refused, so Pilot raises the autoscaler's limits instead.
  5. Someone changes a resource in the console. Pilot sees it, records who, and puts it back because it breaks an objective.

What this is, exactly

Recorded
18 Nov 2026, from a demo workspace on Infrapilot’s own control plane, over the same API the dashboard uses.
The cloud
Simulated. No real AWS resources exist behind it: api (Lambda Function), events (DynamoDB Table), events-wcu-policy (ApplicationAutoScaling ScalingPolicy), events-wcu-target (ApplicationAutoScaling ScalableTarget), orders (SQS Queue), worker (Lambda Function), worker-trigger (Lambda EventSourceMapping).
The decisions
A fixed script standing in for the model, so the demo makes no model calls. In production Pilot’s decisions come from Claude. Everything that checks them is the same code either way: the rules, the bounds, the health checks and the ledger.
The people
dana is the recording script signed in as a person: it approved Pilot’s request through the dashboard’s API after a short wait. The console change is part of the simulated scenario.
The dollars
Estimates from list prices at the traffic of the moment, per month. A ceiling stops Pilot’s own changes from crossing it. When traffic alone pushes an estimate to its ceiling, Pilot is woken to look for savings, which happens here during the order surge.
This page
Plays one static file. It is not connected to a workspace, and nothing here can change anything. Download the recording (JSON, every ledger entry with its hash).