An observability dashboard is now available for developers who build connectors, the MCP-based integrations that bring third-party tools and data into Claude.
The new observability dashboard gives connector developers three core views: adoption metrics (active users, tool calls, directory rank), error and latency diagnostics with per-tool breakdowns, and usage comparisons across Claude, Claude Code, and Cowork.
Developers can now submit connectors directly through a new in-app portal, alongside a blog post detailing the broader connector ecosystem updates.
A year after general availability, bcherny and Cat Wu discuss the evolution of their own usage: preferring auto mode over plan mode, using routines to squash bugs proactively, doing most coding from a phone, and what comes next.
Benchmarks are showing Opus as the top model for long-running autonomous work. Five recommended practices: use auto mode for permissions, use dynamic workflows for orchestration, and three more covered in the linked thread.
The most effective pattern for autonomous work is self-verification plus dynamic workflows. A prompt like "use a workflow to test the result e2e in a browser using Claude in Chrome MCP, especially look for edge cases and UI issues" captures it well.
Good candidates for very long-running Claude sessions include building complex features, cross-language or cross-framework migrations, iterative performance optimization, and hunting down flaky tests.
Dynamic workflows beat static ones on both capability and token efficiency.
Context rot is no longer observed with Opus 4.8, though the team is curious whether other users are still seeing it.
The right frame for evaluating cost is ROI: what would the equivalent work have cost in engineering time? Often the answer is weeks or months.
Skills are designed to be invoked by the model, not directly by users. Just describe what you want and Claude will call the right skills automatically.
Bcherny says he doesn't read the diff until the pull request is up and finalized, a workflow detail that surfaces how he delegates review judgment to Claude Code.
Running /usage gives a breakdown of which skills, MCPs, and plugins are consuming your tokens.
Instead of configuring workflows manually, you can just instruct Claude to use one and it will handle the rest.
Enterprise seat limits are not fixed. Admins can increase them on request.
A previously reported problem appears to be resolved in Opus 4.8, and the team is checking whether users are still encountering it.
Claude Code handles both ad-hoc checks and persistent checks that run on future PRs, depending on what you need.
