Every agent team meets the same wall on the way to production. The demo runs in a small container, so it starts fast. Then the image grows to carry the browser, the SDKs, the internal libraries, and the second model client. The container gets heavier, and the first request of the day starts waiting longer than the work it was asked to do.
On November 12, 2026, AWS moved the fix from an option to the default. The next generation Amazon Bedrock AgentCore runtime is now what a new runtime gets unless you choose otherwise.
Primary source: The next generation AgentCore runtime is now the default (opens in a new tab)
What shipped
AgentCore Runtime V2 reached general availability in September 2026. As of November 12 it is the default platform version, so a new runtime starts on V2 without an opt-in flag.
Two changes carry the release.
Memory works like a page cache. A session starts small, pages in more memory on demand, and returns memory to the platform when the session goes cold. Billing follows what the agent uses rather than the peak it once held.
Cold starts come from a snapshot. AgentCore prepares the environment once, snapshots it, and restores that snapshot for each new instance instead of rebuilding the container every time.
The team's tests put P75 cold starts at about 2 seconds for images from 200 MB up to 2 GB. On the original runtime the same measurement rose from roughly 5.4 seconds to nearly 30 seconds as images grew.
| What you measure | Original runtime | AgentCore Runtime V2 |
|---|---|---|
| P75 cold start, images from 200 MB to 2 GB | roughly 5.4 s, rising to nearly 30 s | about 2 s |
| Session memory | held at peak | paged in on demand, returned when cold |
| How you select it | previously the default | now the default, or set platformVersion to V2 |
| Per-unit rate | lower | higher, against far fewer GB-hours consumed |
Selecting it is one field:
# when you create or update a runtime
platformVersion: "V2"
Two operational details matter. Create and update calls take several minutes while the snapshot is prepared. The container must report healthy on /ping within 120 seconds, so report health only after initialization finishes. AWS also publishes an optimization guide for V2 (opens in a new tab).
AWS notes the per-unit rate is higher on V2 while a session consumes far fewer GB-hours, and it expects most agents to land on a lower bill. Committed baseline pricing is on the way. V2 runs in 10 Regions today.
Why it matters
Cold start time is a governance variable, not only a performance number.
When start time scales with image size, teams strip tools out of the container to keep starts quick. Fewer tools in one image means more calls out to systems outside it, and every call that leaves the runtime leaves the boundary where policy is enforced. Elastic memory and snapshot restore change that trade. A 2 GB image now starts in the same window as a 200 MB one, so the container can carry the tools, the browser, and the model clients it needs, and the policy check stays where the work happens.
Agent request
|
v
[ snapshot restore ] about 2 s at any image size
|
v
+------------------------------------+
| session memory |
| small -> page in -> return on cold |
+------------------------------------+
|
v
[ tool calls: Workspace, Jira, Notion, Slack ]
|
v
[ policy check, approval gate, audit record ]
The second pattern is metering. A runtime that bills on real usage makes cost per task readable at the session level, which is the unit a platform owner can act on. Paired with per-step model routing, agent spend becomes a number you can assign to a team.
How CommandLyne handles this work
CommandLyne runs on a Go-based core on Google Cloud, deployed as Platform Hosted, Single-Tenant Hosted, or Self-Hosted inside the customer's own Google Cloud. Each user gets VM-level isolation, and tool access is scoped per agent rather than per integration. See the platform overview (opens in a new tab).
Agents and workflows follow a Preview, Approve, Adjust, Cancel, Rollback flow, so a fast runtime start still lands inside a path a human signed off on. Team-specific agents reuse that path for repeatable automation across Google Workspace, Jira, Confluence, Notion, Attio, Slack, Discord, Telegram, Salesforce, and browser automation. See agents and workflows (opens in a new tab).
On cost, CommandLyne routes by task complexity, role, latency, and cost across Gemini, Claude, OpenAI, and open models under your own keys, with token-level cost visibility and no reseller markup. That is the layer that pairs with a runtime billed on real usage: the runtime makes the session cheaper to hold, and routing makes each step inside it cheaper to run. The dynamic routing guide (opens in a new tab) covers the routing rules, and governance and multi-model control (opens in a new tab) covers the three-tier hierarchy of Organization Admin, Team Admin, and User.
Ad-hoc versus governed
| Question | Ad-hoc agent stack | Governed agent fleet |
|---|---|---|
| Where the agent runs | one shared host | per-user VM-level isolation |
| What the agent can touch | whatever the service account holds | scoped tool access per agent |
| Who approves a high-stakes step | nobody, after launch | human sign-off checkpoints |
| What a session costs | a monthly total | token-level cost per session and per step |
| Which model runs a step | the one in the code | routed by complexity, role, latency, and cost |
| What you can show afterwards | application logs | a full audit trail |
Where to start
- Set
platformVersiontoV2on a non-critical runtime and compare P75 cold start at your real image size. - Measure GB-hours per session before and after. That number moves faster than the per-unit rate.
- Move one stripped-down tool back into the container and confirm the start time holds.
- Assign a cost owner to one agent, and route its cheap steps to a smaller model.
- Put an approval checkpoint in front of the one step that touches money or production.
Try CommandLyne to run governed agents across your own tools under your own model keys, or book a demo with our team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

