Serverless & AI in brief for week 44 is done, and picking the stories was harder than usual this time. In the week from October 26 to November 1, 2026, GitHub shipped what felt like half a year of features at Universe, while AWS released a whole set of building blocks for Lambda and AI agents. On top of that, Node.js 26 moved into long-term support.
One thread runs through almost every announcement: AI agents get more permissions, and at the same time the guardrails you use to limit those permissions grow too. That is what this review focuses on. For each item you get my take and a link to the original source. Most topics also have a detailed article with a step-by-step guide.
What you will learn in this article
AWS had no single headline this week, but four updates that fit together nicely. Two of them make AI agents on AWS more tangible, and two improve how you run Lambda functions.
Since October 29, you can expose Step Functions workflows as tools for AI agents through Amazon Bedrock AgentCore Gateway. The gateway speaks MCP (Model Context Protocol), the open standard agents use to discover and call tools. The agent decides on its own when to start a workflow and with which input. It can also query executions, for example which runs failed and at which step. This works with both Standard and Express workflows, and you pay the regular prices for Step Functions and AgentCore Gateway. One day earlier, AWS released version 1.0 of the AgentCore CLI. It starts faster, has consistent commands and offers agentcore migrate as a path for existing 0.x projects.
For me, this is the most important AWS announcement of the week. The model decides what should happen, and the workflow defines how it happens, step by step and traceable. That separation makes agents much easier to justify inside a company. I show how to set this up with the AWS CDK in Step Functions agent tools with AgentCore Gateway, and how to get started with the CLI in AgentCore CLI 1.0: deploying an AI agent with TypeScript. On October 28, AWS released the Durable Execution Telemetry Exporter, a plugin for Lambda durable functions. These are functions that model long-running processes with waiting periods across several invocations. The exporter sends the history of every execution to a destination of your choice: status, duration of each operation, step inputs and outputs, and errors. You can redact sensitive fields and adjust the sampling rate. AWS lists 13 built-in destinations, including CloudWatch Logs, S3 for queries with Athena, DynamoDB, Aurora and OpenSearch. TypeScript, Python, Java and Go are supported, and the plugin costs nothing extra.
So far, this was exactly the weak spot of durable functions: a process spanning two days fell apart into many log lines from many invocations. Now you get one record per execution, and questions like “What has been stuck in step X since yesterday?” turn into a simple query. You can find an example with DynamoDB as the destination in Monitoring durable functions: telemetry with TypeScript.
Also on October 28, AWS introduced SOCI (Seekable OCI) for container images on Lambda Managed Instances. Instead of down the whole image before starting, Lambda fetches only the files needed in each startup phase. AWS mentions up to 70 percent shorter start times. To use it, you create a SOCI index for the image, push it to Amazon ECR, point the function at the indexed tag and publish a new version. Lambda detects everything else on its own.
Keep the limitation in mind: this only applies to Lambda Managed Instances, not to classic Lambda functions. It becomes interesting mainly for large images with ML libraries or models. For a slim Node.js image, I would skip the extra build step. How to measure the effect before you commit is covered in SOCI lazy : faster Lambda container starts.
Since October 28, the Identity Store of IAM Identity Center sends events to Amazon EventBridge. They fire when users, groups, group memberships or the identity store itself are created, updated or deleted. Additional costs only come from regular EventBridge usage, and the feature is available in all Regions with Identity Center.
That sounds unspectacular, but it saves a lot of tinkering. If you used to poll group changes with a nightly script, you can now react immediately, for example during offboarding. Access that lingers for days after someone changes teams shows up in almost every audit. I would still keep a nightly reconciliation, because events can get lost or be processed incorrectly. I describe the complete solution with the Serverless Framework in Identity Center EventBridge: automating offboarding.
On October 28, GitHub published more than 30 changelog entries for Universe. Two topics among them affect teams directly, because they change who decides what in a repository.
Already on October 26, assisted permissions became generally available in the Copilot CLI, the Copilot app and VS Code. A language model checks every request from the agent to run a command, access a path or open a URL. Whatever it classifies as clearly safe runs automatically, everything else waits for you. Administrators can, for example, block “Allow All” while still allowing the assisted mode. Two days later, Copilot approvals in code review followed: once the feature is enabled, an approval from Copilot counts toward the required approvals just like a human one. It is off by default, and repository admins decide for which paths Copilot may approve. There is also a new review level called “Max” in public preview, where several review agents using models from different providers check in parallel and another agent merges their findings. And GitHub Enterprise now includes Copilot for every licensed user, but without an included allowance. Every bit of AI usage is billed against the company budget from the first request.
This is where I would slow down before anyone flips the switch. An approval in review is more than bug hunting, it spreads knowledge and responsibility across the team. Clear rules about which paths the AI may approve on its own are therefore a team decision, not a settings question. You can find a traffic light model for this in AI code review in your team: rules for Copilot approvals, and the matching allow, ask and deny rules in Setting up Copilot assisted permissions safely.
Lockfiles for GitHub Actions have also been generally available since October 28. The file .github/workflows/actions.lock records which code is behind each action. You create it with the gh actions-lock extension for the GitHub CLI, and with the “Require lockfile” policy, admins can make it mandatory at enterprise, organization or repository level. GitHub itself points out that a lockfile says nothing about whether a dependency is safe. The second big Actions announcement: GitHub Agentic Workflows are now generally available. You describe an automation such as issue triage or analyzing failed CI runs in a Markdown file, and the tool translates it into regular Actions YAML. By default, agents run with read-only permissions in an isolated environment, and their results are checked before they are allowed to change anything.
For me, both belong together. The more agents work in pipelines, the more important it becomes that the pipeline itself cannot be swapped out silently. My advice: introduce the lockfile first, then let agents into workflows. With gh aw forecast you can estimate the AI costs of an agentic workflow in advance, which I would do before the first run. I explain how to introduce the lockfile step by step in GitHub Actions lockfile: pinning workflows safely.
With version 26.13.0, Node.js 26 moved into long-term support on October 28, codename “Lithium”. Updates are planned until the end of April 2029. In terms of content, the release differs from 26.12.1 only in its metadata. If you are coming from Node.js 22 or 24, though, you should know one thing: since Node.js 25, the official Linux binaries are linked against libatomic, so depending on your distribution you may need to install that package. On AWS Lambda, the nodejs26.x runtime is available as a public preview, and according to the documentation general availability is planned for November.
For me, this is the right moment to write the upgrade plan instead of waiting for the deprecation email. Test your functions against the preview now, so that at GA you only need to switch. What changes and how to check it with TypeScript and CDK is covered in Node.js 26 LTS: planning your Lambda upgrade with TypeScript. This week the tools noticeably give AI agents more room. Copilot may run commands and approve pull requests, agents start workflows on AWS and work inside GitHub Actions. At the same time, we get lockfiles, path rules, read-only permissions and sandboxes. What matters most, therefore, is where the limits of agents are and who sets them.
If you only have one hour for this week’s news, this is how I would spend it:
nodejs26.x preview runtime.
The next review will be published on Monday, November 9, and covers week 45. You can find the first issue on week 43 and all upcoming ones on the page Series: Serverless & AI in brief. If an important piece of news is missing, feel free to add it in the comments.
Now it’s your turn: would you let Copilot approve pull requests in your team, and if so, for which files?
This post was created with AI assistance.
👉 Next step: Working with AWS CDK? Grab my printable AWS CDK cheat sheet with commands, concepts and a stack skeleton.
Originally published on sascha-ms.de. More articles on serverless, AI in everyday development and leading tech teams: sascha-ms.de/en