Claude Opus 5: Long-Running Agents That Actually Ship

Anthropic's Claude Opus 5 is now available in Microsoft Foundry. The headline claim is that it can keep working for hours—overnight, even—without collapsing mid-task. For anyone who has handed a multi-step refactor to an AI and gotten back a "sorry, I can't complete that" after two files, this matters.

The real shift: codebase reasoning, not code writing

Writing a decent function isn't the bottleneck anymore. The hard part is understanding how a change ripples through a repo with 15 dependencies, three architecture decisions from 2021, and a config file that someone broke a year ago. Opus 5 is tuned for exactly that: navigating a large codebase with the judgment of a senior engineer who reads the whole diff before touching a line. It adjusts its approach when a project evolves mid-task and recovers from errors instead of hanging up. That behavior saves you a morning of debugging why the linter suddenly hates your PR.

Long-running agents that don't lose the plot

Most enterprise workflows aren't a single prompt away. Research, document review, operational automation—these run across hours or days. Opus 5 is built to plan a workflow, hold context across tools and subagents, and adapt when something breaks. This matters if you've tried to automate a process involving three different APIs, a PDF extraction step, and a spreadsheet output. The model can run end-to-end, produce professional-grade slides or docs, and stay in scope without wandering into unrelated Kubernetes best practices. However, reliably evaluating performance in these long-running tasks can be difficult—these models are not easily caught by standard measures, and simple quota-based checks may not suffice.

Vision that actually sees, not just reads

Hardware manuals, scanned invoices, dashboards, screenshots of internal tools that never got proper docs—critical information lives in images. Opus 5 brings advanced vision to the Opus line, and crucially it reasons over visual content, not just extracts text from it. That means it can look at a chart in a dense financial filing and understand what's happening, not just OCR the axes. Combined with Foundry's IQ feature that connects agents to your existing documentation and process knowledge, you get something close to a junior analyst who actually reads the room.

Computer use that doesn't feel like a party trick

The model can also navigate software interfaces and complete multi-step tasks across applications. For automating repetitive business operations, this is practical: it's about exercising judgment when the UI changes or a popup appears, not just clicking buttons.

Platform matters, too

A powerful model alone won't get you past compliance review. Foundry provides evaluation tools, Azure-native auth and governance, grounding with enterprise data, and continuous monitoring. For teams shipping code tomorrow, that means you can put this into production without a six-month security audit detour.

Claude Opus 5 is available now in the Foundry model catalog. If your workflows have been waiting for an AI that works through coffee breaks instead of taking them, this is worth evaluating.