A human-curated opener makes AI-written PRs easier to review
· Summary by Vibe Code Weekly
Simon Willison says agent-written pull request descriptions can be accurate but overwhelming. He now leads with a short, human-approved explanation and puts the longer agent-generated details inside collapsible <details><summary> markup, which GitHub Markdown supports.
Takeaway A small formatting change can preserve an agent’s full PR documentation while making the key context quicker for reviewers to grasp.
Replit Agent can now prepare one-time supply purchases through Link and Visa, stopping for the owner’s approval before payment. It can also create and manage Google Ads campaigns, drafting assets, recommending budgets, and adding conversion tracking; purchasing is available to Core and Pro users in the US, while ads are available to Pro users worldwide.
Takeaway Replit is expanding its coding agent into business operations, letting founders handle purchases and customer acquisition from within Replit.
Simon Willison used Claude Code to build Beep Locator, a Python web app that coordinates phones, tablets and computers to listen for beeps and triangulate their source. The system located an intermittent beep to within about 40 cm, leading him to an old carbon monoxide detector with dead batteries.
Takeaway The project shows how a coding agent can turn acoustic-localization research into a practical tool for a real-world problem.
Datasette Agent 0.5a0 introduces a panel that keeps its agent chat alongside Datasette pages, such as tables and queries. Paired with datasette-apps, it lets users create database-backed apps from prompts and then refine them through the adjacent chat.
Takeaway The release brings an agent-led build-and-edit workflow directly into database-backed apps.
Replit describes Agent Habitat, an environment for giving company agents access to internal knowledge and systems within organization-defined boundaries. Its controls include scoped permissions, approval rules, credentials kept outside the agent’s context and added to authorized requests by a broker, plus isolation between the agent and the code-execution sandbox.
Takeaway The architecture shows how teams can delegate more work to coding agents while limiting access, protecting credentials and keeping activity accountable.
Replit says Cloud Agents let non-engineers describe changes in plain language, preview them in a running app, and open pull or merge requests for engineers to review. Agents can test changes in cloud sandboxes, attach screenshots and video, and carry review feedback forward; the offering is available to Pro and Enterprise customers, with workspace-memory improvements in early access.
Takeaway Replit is extending coding agents from app creation into company software workflows, while keeping engineers in charge of what ships.
Replit’s new Monitoring Agent uses production logs and metrics to investigate alerts, identify causes, and validate fixes. Builders can enable automatic publishing so the agent deploys validated fixes and emails a report; publishing remains optional.
Takeaway Replit is extending its coding agent into production maintenance, offering small teams an automated on-call option for apps hosted on the platform.