AWS Cuts AI Agent Startup Time Up to 15x and Memory Per Session by Up to 68%
An agent needs its own isolated sandbox, and until now teams had to choose: keep ready machines idle and pay for downtime, or wait through a cold start. In the updated AgentCore runtime from AWS, a session starts in under 2 seconds, up to 15x faster than the previous version, and memory per session for agents with pauses drops by up to 68%. As a result, warm pools are no longer needed, and agent budgets are worth recalculating.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Picture an agent that fixes bugs in your repository: it runs code, executes tests, and browses websites. You can't let such an agent into a shared environment; it needs its own isolated sandbox. And then the trade-off begins: if you keep ready sandboxes on standby, you pay for idle time, and if you create them on demand, the user waits through a cold start.
AWS has updated AgentCore runtime, the managed compute layer within Amazon Bedrock AgentCore, and promises to remove that choice. An agent session now starts in under 2 seconds, which AWS estimates is up to 15x faster than the previous version. For agents that pause between steps, memory use per session falls by up to 68%, because an idle agent no longer holds memory for nothing.
What the problem was
Agents have long moved beyond single prompts: they run multi-step tasks for hours, execute code, work with sensitive data, and act on behalf of users. That is why every session needs isolation; otherwise, one user's code could reach another user's data. When there are thousands of agents, homegrown sandboxes become complex, expensive, and unpredictable.
AWS cites industry surveys. According to Deloitte, only 15% of organizations have scaled orchestrated multi-agent systems, and 70% admit they cannot yet trust their agents or manage them. According to McKinsey, 40% of large companies are already scaling AI agents, up from 27% a year earlier. AWS does not disclose the survey methodology, but the conclusion is clear: many teams are stuck not on model capabilities but on launching agents securely and cheaply.
What changed
The core guarantee is unchanged: each session runs in its own hardware-isolated Firecracker microVM (a lightweight virtual machine that boots in fractions of a second), with its own identity, access policy, and audit log. On top of that, there are four changes.
- Startup and size. A session starts in under 2 seconds regardless of image size or the number of parallel launches. Limits quadrupled: up to 8 vCPUs (virtual processor cores) and 32 GB of memory, and container images can weigh up to 10 GB. That is enough for building large codebases, browser-based workflows, and running models locally. According to AWS, warm pools and pre-provisioned clusters are no longer needed.
- Pay for work. An agent spends most of its time waiting for a model, a tool, or a human to respond. Now the environment frees memory within a few minutes after the agent releases it, and an idle agent scales to zero. For conversational agents with pauses, this is what delivers up to 68% memory savings per session.
- Plan for persistent agents. For agents that run continuously, there is a committed baseline: you fix the minimum memory per session and pay for it as a minimum. Beyond that, you can temporarily draw up to four additional such volumes on demand. The rates are tied to AWS Lambda MicroVM pricing.
- Data isolation. Previously, an agent sat in a secured session while the files it needed lived outside. Now a session can dynamically mount a file system, for example an Amazon S3 bucket via FUSE (which makes storage look like an ordinary folder), and save results there. Data stays within the trust boundary, and access rules are set once.
What customers say and what to watch
AWS shares customer feedback. Coinbase reports that cold session starts dropped from nearly a minute to under 5 seconds even at peak, and its weekly volume of agent launches more than doubled. Mission Cloud saw 35% faster starts for inference requests, 77% faster health-check starts, and 17% lower cost per call on the same workload. Zuora reduced memory per call by roughly 60% in many scenarios while keeping latency stable. Okta uses the combination with Lambda MicroVM so that agents can check code, run tests, and open pull requests in disposable sandboxes.
The caveats are worth keeping in mind. The figures either read as marketing "up to" numbers or come from the customers' own words in the press release; there are no independent measurements. The gains depend on the workload profile: if an agent keeps its memory occupied the whole time, the savings will be smaller than 68%. AWS only promises to bind each session to its own storage folder via AWS session credentials by re:Invent, so data isolation is not yet complete.
The practical takeaway for teams building agents: if you keep warm pools or homegrown sandboxes for fast starts, compare their cost with the new pay-for-actual-work model. It is easiest to start with agents that have many pauses between steps, where memory savings are greatest. The service is already available in the Americas, Europe, and Asia Pacific regions.
- Nov 21, 2026AWS and NVIDIA show how to split a coding agent's requests between Claude Opus 5.5 and Nemotron
- Nov 21, 2026AgentCore V2 from AWS launches agents from memory snapshots, but for chat and voice even 2 seconds is too long
- Nov 21, 2026You can cut your Amazon Bedrock bill without switching models: Flex costs half as much as Standard
- Nov 21, 2026Natera's AI Agent Cuts Instrument Failure Triage from 17.4 Hours to 8 Minutes
Key Questions Answered
What the problem was?
Agents have long moved beyond single prompts: they run multi-step tasks for hours, execute code, work with sensitive data, and act on behalf of users. That is why every session needs isolation; otherwise, one user's code could reach another user's data. When there are thousands of agents, homegrown sandboxes become complex, expensive, and unpredictable. AWS cites industry surveys.
What changed?
The core guarantee is unchanged: each session runs in its own hardware-isolated Firecracker microVM (a lightweight virtual machine that boots in fractions of a second), with its own identity, access policy, and audit log. On top of that, there are four changes. Startup and size. A session starts in under 2 seconds regardless of image size or the number of parallel launches. Limits quadrupled: up to 8 vCPUs (virtual processor cores) and 32 GB of memory, and container images can weigh up to 10 GB.
What customers say and what to watch?
AWS shares customer feedback. Coinbase reports that cold session starts dropped from nearly a minute to under 5 seconds even at peak, and its weekly volume of agent launches more than doubled. Mission Cloud saw 35% faster starts for inference requests, 77% faster health-check starts, and 17% lower cost per call on the same workload. Zuora reduced memory per call by roughly 60% in many scenarios while keeping latency stable.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.