A faster session start changes where an agent team should do work. When AWS restores a prepared runtime instead of initializing a container for each new session, the question becomes which state can safely be inherited. Getting that division right can save real time. Getting it wrong can make a session begin with stale credentials or outdated application assumptions.
On November 12, AWS announced that AgentCore Runtime V2 had become the default for newly created runtimes in listed regions. This is a change to the managed compute layer, not a new model or a claim about the separate Bedrock Managed Agents preview. It also differs from the connection renewal discussed in our Nova Sonic and Strands Bidi assessment: V2’s snapshot starts a new execution session from prepared state.
What AWS changed
AWS says V2 restores an initialized snapshot into a separate Firecracker microVM for each session. For a 2 GB container image it reports a P75 startup below two seconds, versus 29.7 seconds on the previous generation. It lists capacity up to 8 vCPUs, 32 GB of memory and 10 GB images, and reports up to 68% less billed memory per session for multi-turn agents. These are service limits and vendor measurements. They do not predict time to the first useful answer or the total bill. November 12 service announcement
The dated announcement says existing runtimes keep their version unless changed, and lists the regions where V2 is now the creation default. The microVM developer guide still says omitting platformVersion at creation selects V1 and shows a narrower region list. The release notes separately record an expansion to nine more regions. These public sources disagree about the omission behavior. For a migration, pin the intended version explicitly where the deployment interface supports it, then inspect the resulting runtime before routing traffic. We did not resolve the discrepancy with a control-plane call.
Separate inherited preparation from session state
The November 13 architecture guide distinguishes memory that survives restoration, files read on demand, and network connections that must be reestablished. Reading a file during initialization does not by itself make reads faster in a restored session. Keeping reusable data in memory, or using a build hook for assets the workload needs early, changes what is ready at first use. AWS startup architecture guide
For an illustrative document agent, separate the immutable parser and pinned lookup tables from mutable policy, tenant data and in-flight work. The first group is a candidate for preloading. The second needs to be fetched or validated against the actual invocation. If a policy table was safe at image build time but gets revised independently, treating it as static changes the meaning of later answers even when the system starts faster.
Write an inventory of what initialization holds: shared configuration, reference data, SDK clients, open connections, test fixtures and deferred work. For each item, record when it becomes valid and what should invalidate it. Test a newly restored session immediately after rotating the backing configuration. A passing startup probe only establishes service readiness under that probe; it doesn’t prove that the request will use current data or that a dependent service will accept a connection.
A session-specific client has different requirements
AWS’s startup guide encourages creating some clients early to avoid repeating setup. Its session guide describes an important exception: an AWS SDK client created before the snapshot can retain pre-session credentials, which lack the session-specific IAM context key. Where resource policy relies on that key, the guide directs operators to create the client inside the invocation handler so it loads the session’s credentials. It also says the context keys attest to the credentials’ origin and do not prove physical execution inside the microVM. Isolated sessions and IAM context
A separate November 12 AWS engineering post describes storage-folder scoping as a milestone by re:Invent. That future-tense statement and the living guide’s documented credential mechanism should be treated as distinct evidence. Neither proves that a particular mounted volume, identity route, SDK or region currently enforces the intended scope. AWS engineering announcement
For a pilot, use synthetic objects scoped to two authorized test sessions. Verify both allowed reads and denied reads, and record which identity the storage service observed. An empty response can mean denial, a missing object, or an initialization error; classify and retain those results separately. The isolation evidence is useful only when the allowed operation also works and the operator can explain which condition denied its sibling.
Measure readiness and useful work separately
V2 shifts some work from each session to deployment. AWS documents snapshot preparation before a runtime becomes ready, and says a replaced snapshot can remain while sessions already using it finish. Snapshot lifecycle This makes it possible for a version switch to change new sessions while old work still exists. Tie observations to the version that produced them and define what a release rollback should do for ongoing tasks, not just for future requests.
I would run a paired pilot with the same image, pinned model, fixtures and acceptance rubric. Use explicit V1 and V2 configurations in a region where both target configurations are supported. Include short tasks from idle, sustained sessions with pauses, a burst of fresh sessions, and a deployment while existing sessions are active.
| Question | Proposed measurement |
|---|---|
| Did a session start sooner? | Cold-start quantiles and warm-session results, reported separately |
| Did the user receive useful work sooner? | Time to first validated artifact and time to accepted completion |
| Was the restored application current? | Versions of policy, data and dependent services observed after restore |
| Did migration preserve isolation? | Allowed and denied session-scoped fixture operations |
| Did workload costs fall? | Billed memory and CPU, model/tool charges, rejected runs and operator effort |
Record readiness time for each deployment outside the request timing. Attribute time spent reading restored assets, reestablishing connections and authorizing the actual operation to the task itself. A shorter runtime startup may be offset by a slow first read or an external wait. Plot cold and warm calls separately so changing the mix cannot create a false speedup.
Evaluate the bill under the way the agent works
An agent that releases working memory after each burst has a different footprint from one that retains a full document index during long pauses. More efficient metering can help the first workload without automatically improving the second. Provisioned commitments, idle time, storage, external tool costs and model calls all influence the decision. Compare observed billed resources against timestamps and accepted outcomes, especially if the model is the main cost driver.
My adoption test would be lower total cost or faster accepted work under stable isolation and correctness, including deploy and rollback effort. A synthetic cold-start win is enough to justify testing that hypothesis. Repeated task acceptance is what would justify moving a workload.