It was a busy week for AI agents. Most of the big moves came out of Microsoft Build 2026, but there were a few other stories worth your attention too. Here's what mattered.
GitHub Copilot is now its own app — and agents are the reason why
GitHub launched a dedicated desktop app for managing parallel AI agents, built around a "My Work" dashboard, isolated Git worktrees, and a new concept called Canvases — bidirectional work surfaces where you can inspect and steer agent work as it happens. The Copilot SDK is now GA across Node.js, Python, Go, .NET, Rust, and Java. This is GitHub's move to own the agent orchestration layer.
Read more
Microsoft wants Windows to be the enforcement layer for AI agents
Microsoft Execution Containers (MXC) is a new policy-driven, OS-level containment model for AI agents running on Windows and WSL. It's not a product — it's a primitive. You declare what an agent can access; the kernel enforces it. Integration with Defender, Entra, Intune, and Purview is coming in preview in July. If you're building or deploying local agents in enterprise environments, this one's worth understanding now.
Read more
AI agents are now living in your Windows Terminal
Microsoft also announced AI agents embedded directly in Windows Terminal, which means agent interactions are moving out of IDE plugins and into the command line itself. The implications for how developers actually work day-to-day are significant.
Read more
Claude Code can now handle workflows that were too big to automate
Anthropic's Claude Code added dynamic workflow capabilities aimed at multi-step tasks that previously required too much human intervention to automate reliably. It also now catches security vulnerabilities inline while you write — not after a separate scan. Two separate but related moves toward making the coding agent more useful for real production work.
Read more — workflows | Read more — security
Your test and dev data is a bigger risk than you think
Around 80% of organizations are still passing production data into non-production environments. That was a manageable risk a few years ago. Now AI can cross-reference scattered data points and surface exposures no human analyst would catch. The fix isn't new tooling — it's masking data before it leaves production and making compliant datasets self-service so developers don't work around the process.
Read more
Most organizations are flying blind on AI spend
AI costs are scaling faster than the visibility into where they're going. This piece looks at what good AI cost management actually looks like and how to get there before the bills become a problem.
Read more
Snowflake Summit 26: What developers need to know
If you missed Snowflake Summit, here's the short version of what's relevant for developers, engineers, and architects.
Read more
The through-line this week: agent autonomy is expanding fast, and the infrastructure to contain, govern, and observe it is finally starting to catch up.
See you next Friday.
Developer Weekly Briefing is published every Friday on Coder Legion. Written by Tom Smith.