#30 - GitHub agents, Cursor 2.0 reviews, and more 🚀

Plus: why devs trust AI code too much

November 04, 2025

Hey readers! 🚀

This week brings some fascinating shifts in how we're building software. From GitHub opening its doors to competing AI agents, to new research showing we might be trusting AI suggestions a bit too readily, there's plenty to unpack. Let's dive in!

This Week's Highlights 🎯

GitHub Opens the Gates with Agent HQ

GitHub Embraces the Coding Agent Competition With Agent HQ marks a significant strategic shift for the platform. Rather than forcing developers into a single AI ecosystem, GitHub is building infrastructure that welcomes competing agents from Anthropic, Google, OpenAI, and others into a unified interface. – Frederic Lardinois

The new Agent HQ platform includes Mission Control, a central dashboard accessible via web, VS Code, mobile, and CLI where developers can monitor and steer multiple AI agents working on the same codebase. What's particularly interesting is GitHub's embrace of competition rather than vendor lock-in. As COO Kyle Daigle explains, "developers are going to choose the tools they want to choose, and they're rarely working 100% alone."

"Our goal with Agent HQ is that we have a single place where you can use basically any coding agent that wants to integrate, and have a single pane of glass — a mission control interface, where I can see all the tasks, what they're doing, what state of code they're in."

The platform also introduces GitHub Code Quality for org-wide visibility into maintainability and reliability, including a first-line review step specifically for AI-generated code. Enterprise features like audit logging, security policies, and access controls address the governance concerns that have slowed AI adoption in larger organizations.

The Trust Problem: When Developers Stop Questioning AI

Software developers show less constructive skepticism when using AI assistants presents research from Saarland University that should give us pause. The study found that developers working with GitHub Copilot scrutinize AI-generated code less critically than they would a human colleague's work, leading to reduced knowledge transfer and potential technical debt accumulation. – Saarland University

The research compared human-human pair programming with human-AI pairings and found that AI-assisted teams had narrower, more code-focused exchanges. Developers were more likely to accept AI suggestions without critical evaluation, assuming the code would work as intended. This matters because the rich, questioning dialogue that happens in human collaboration is where much of the learning and quality improvement occurs.

"Developers tend to scrutinize AI-generated code less critically and they learn less from it."

The researchers conclude that AI assistants excel at simple, repetitive tasks but can't yet replicate the depth of human collaboration. This suggests we need to be more intentional about maintaining critical thinking habits even when AI makes development feel effortless.

Cursor 2.0 Doubles Down on Multi-Agent Development

Cursor 2.0 pivots to multi-agent AI coding, debuts Composer model introduces a new in-house model designed specifically for low-latency coding workflows. Composer completes most tasks in under 30 seconds and is optimized to navigate large projects while tracking dependencies across multiple files. – Ryan Daws

The real innovation is Cursor's multi-agent architecture, which coordinates several AI agents working in parallel on isolated git worktrees. This allows multiple agents to tackle different aspects of the same problem simultaneously, then select the best solution. Early testers report that this approach significantly improves output quality for complex tasks.

Cursor 2.0 Expands Composer Capabilities for Context-Aware Development adds that Composer builds contextual awareness over time, remembering past edits and design patterns to provide more consistent suggestions. The platform also includes a built-in browser with Chrome DevTools integration, allowing developers to pinpoint UI issues and send exact HTML elements directly to the AI chat. – InfoQ

Terminal AI: A Different Approach to Context Management

You've Been Using AI the Hard Way (Use This Instead) makes a compelling case for moving AI interactions from the browser to the terminal. NetworkChuck demonstrates how tools like Gemini CLI, Claude Code, and OpenCode enable project-aware context management through local markdown files that persist across sessions and sync between multiple AI tools. – NetworkChuck

The key insight is ownership: when your AI context lives in local files rather than scattered browser chats, you can version control it with Git, copy it anywhere, and avoid vendor lock-in. NetworkChuck shows how he uses multiple AI agents for research, critique, and content creation, all working from the same synchronized context files.

"Everything I'm doing talking with these three different AIs on a project, it's not tied in a browser, it's not tied in a GUI. It's just this folder right here on my hard drive."

The Evolving Landscape of AI Development Tools

Last Month on the Landscape: 8 Tools in the Spotlight surveys the current ecosystem, highlighting tools that represent different integration philosophies. Shell GPT and Warp meet developers where they already work, while Kiro and Agents.md propose more fundamental workflow changes through specification-driven development. – AI Native Dev

The article notes a pattern: successful tools either enhance existing workflows with minimal disruption or offer compelling enough benefits to justify switching. Taskmaster's orchestration approach for Claude and DeepWiki's knowledge management focus show how the ecosystem is addressing specific pain points rather than trying to replace entire development environments.

Security and Ethics in AI-Assisted Development

How Software Development Teams Can Securely and Ethically Deploy AI Tools emphasizes that even the best LLMs generate incorrect or vulnerable code nearly two-thirds of the time. The article warns against using one AI to generate code and another to review it with minimal human oversight, calling this a false sense of security. – SecurityWeek

Beyond security, ethical and legal concerns around copyright and training-data provenance require careful governance. The piece recommends establishing internal guidelines, maintaining traceability, monitoring for unapproved tools, and upskilling developers to recognize vulnerabilities in AI-generated code.

Academic Perspectives on AI in Software Engineering

Generative AI in Software Engineering: Transforming the Software Development Process provides a comprehensive white paper from Accenture and DFKI examining GenAI's impact across the entire software lifecycle. The research emphasizes human-AI collaboration, where AI handles formalized, repetitive tasks while humans focus on problem formulation and validation. – Accenture & DFKI

"The collaboration between human software engineer and AI tools is at the core of effective integration of GenAI into SE. AI can take over formalized, technical, repetitive, error-prone, and effort-intensive tasks."

Meanwhile, CMU Students Put AI Coding Tools to the Test describes Carnegie Mellon's experimental course where students build software without writing code themselves, instead learning to instruct, evaluate, and correct AI-generated code. The course is a deliberate pedagogical experiment to rapidly learn how to teach AI-assisted development rather than waiting for long-term consensus. – Carnegie Mellon University

Quick Hits 📰

Made with ❤️ by Data Drift Press

Have thoughts on this week's stories? Hit reply – I'd love to hear from you!