Last Week in AI: 3rd Edition

Hey all, and welcome back to Last Week in AI, your weekly roundup of what engineers, builders, and hiring leaders need to know.

This week brought meaningful updates across the developer stack. Cursor and GitHub rolled out new ways to coordinate multiple agents. OpenAI introduced a tool to automate security review. Chrome DevTools added model-powered performance analysis. Claude’s memory feature started rolling out to individual users.

More of the workflow is being shaped, executed, and reviewed by systems that retain context and operate across scopes.

Here’s what shipped:

1 | Cursor 2.0 introduces multi-agent worktrees and a new foundation model

Cursor now lets you run up to eight AI agents in isolated Git worktrees, each capable of modifying, documenting, or refactoring different parts of the same project. You can assign agents by scope or function and compare their approaches side by side.

The new Cursor-native model is tuned for multi-turn, cross-file reasoning. It handles long-range structure and decisions better than off-the-shelf models. Rob Shocks shared a good breakdown, here.

This release pushes AI into the role of system participant, not sidecar. We’re seeing the developer become a reviewer, editor, and coordinator (more oversight, more decisions, more architectural thinking across moving parts).

2 | GitHub launches Agent HQ and Plan Mode across ecosystem

Agent HQ gives teams visibility and control over how copilots operate across GitHub, VS Code, and CLI. You can configure agent roles, enforce file access boundaries, and see what each copilot is doing in real time.

Plan Mode lets Copilot generate multi-step execution outlines before writing any code, giving you a way to shape intent upfront.

These features are evidence of GitHub’s bet on agent management as a core skill for engineering teams. Planning, orchestration, and oversight are no longer optional!

3 | OpenAI launches Aardvark for continuous code security

Aardvark was just launched (in Beta). It reads your repo with long-context attention, flags security issues with ranked severity, and drafts precise inline patches. It explains exploit paths, validates bugs in a sandbox, and suggests one-click fixes using Codex.

In scanning Reddit and X, sentiment seems mixed. Some engineers see this as a real productivity unlock, while others flagged the absence of public benchmarks.

Regardless, Aardvark won’t replace review; it’ll enable teams to surface and triage security risks before they spread inside the environments they already use.

4 | Chrome DevTools adds Gemini-powered trace analysis

Chrome 142 embeds Gemini into performance tooling. You can ask direct questions about frame drops, scripting delays, or paint costs, and get actionable answers grounded in trace timelines.

This creates a new debugging pattern where understanding becomes conversational and explanations are contextual.

Frontend teams will ship faster with fewer handoffs to performance specialists. As context-aware debugging gets built into everyday tools, developers who are able to reason through performance without switching roles get a step ahead.

5 | Claude memory rolls out across Pro and Max plans

Claude’s memory feature is now live for Max users and rolling out to Pro subscribers over the next couple weeks. It lets the assistant retain preferences, architecture patterns, and key project context across sessions. 

Users can see, edit, or delete memories directly, and even create separate memory spaces to isolate work streams. This makes Claude more usable for longer-lived or multi-track work. 

The response has been split. Some developers appreciate the continuity, while others report confusion or hallucinations from misapplied context. 

Teams using memory well will treat it like versioned knowledge… not ambient magic. That skill (making tool memory an asset, not a liability) is becoming part of how good engineers work across time and scope.

The through-line

More of the work is happening before any code gets written. Tools are planning tasks, assigning scopes, and drafting outputs. What matters is how clearly the goals are set and how well the structure holds up.

The teams making the most progress are the ones spending time upfront getting specific about what’s being built and how.

What’s new at HackerRank

A few weeks back, we hosted our Customer Advisory Board in NYC with engineering leaders to trade notes on how AI is reshaping developer skills, workflows, and evaluation. The feedback is directly feeding into how we design HackerRank for next-gen developers.

We learned that the number of candidates using AI assistants in HackerRank assessments has exploded since June, now scaling by the tens of thousands weekly as AI-native testing becomes the norm (see Vivek’s tweet): 

This surge shows how quickly AI-native workflows are becoming standard. (Hiring leads reading: If your interview process doesn’t reflect how developers build with AI, you’re likely filtering out some of the strongest talent.)

Just Released

• Last week’s AI roundup on Hackonomics →
• Handpicked Hackonomics drop: Train Your Own ChatGPT From Scratch (for $100) →
• New blog post: The AI Researcher Arms Race: Inside Tech’s Priciest Talent War  →
• New blog post: Inside the Rise of the AI Infrastructure Engineer →

The future is shipping faster, and the next-gen developer is already debugging it! That’s it for this week. Back next week with more AI news.

In case you missed it: Edition #1 (10/9) | Edition #2 (10/24)