Background startup was throttling timers and dropping the JS engine into power-saving mode. The fix boots at full speed even before the window is visible, along with additional smaller performance improvements.
Bcherny amplified the ClaudeDevs startup speed announcement, noting that slow boot makes the app feel sluggish for daily users and that further improvements are still in progress.
A user reported the desktop app using 3x more RAM than a month ago, forcing them to fall back to the terminal. bcherny said the team has been shipping performance wins daily for about a week, with more on the way.
A user celebrated the recently shipped startup-speed improvement after Lydia Hallie had previewed it in comments. She replied confirming more improvements are coming.
Auto mode in Claude Code uses a separate classifier model that reads plain-English rules to decide what actions to allow or block. Adding "$defaults" to your config preserves the built-in rules while layering on your own.
Running `claude auto-mode critique` will review your custom auto mode rules and flag anything ambiguous, redundant, or likely to cause false positives, helping you tighten up plain-English permission configs before they cause problems.
Responding to criticism that Anthropic stays silent on major issues like outages, guardrails complaints, and concerns about Opus 5 while posting frequently about minor updates, bcherny said the team will work to change the dynamic and is actively improving all of the above.
A user described Opus 5's behavior and Fable's over-refusals as an existential crisis for Anthropic, despite being a heavy token user and strong supporter. bcherny responded simply that the team is working on improving all of these.
Responding to a request for Cursor or Codex-level UI polish in Claude Code, bcherny said quality has improved a lot recently with more improvements on the way, and invited users to tag him directly when they spot issues.
Responding to a user complaint that Anthropic isn't as responsive to customers as it should be, bcherny opened the floor by asking directly what could be improved.
Responding to feedback comparing Claude to Codex's unified task interface, bcherny confirmed that session-to-session communication and starting local sessions from mobile already work today, with more features coming soon.
Throughout its history, software development has been unreliable: late, over budget, and missing user needs. AI-powered "software factories" represent a genuine promise to fix that, especially for SMBs that could never afford quality custom software.
Companies whose core competency isn't software stand to benefit most from AI-driven "software factories" that make development reliable and predictable. But building genuinely new software products will remain risky and unpredictable, just profitably so.
Trq212 argued that most MCPs and CLIs are intentionally limited because companies fear agents cutting them out of the loop. The solution is to make them fully featured but structure pricing so the company captures more value from agent usage.
The Claude desktop app is built on Electron, which the team has gotten feeling "really good." Native macOS prototypes exist but come with downsides, so the plan is to keep pushing Electron first and only reach for native if necessary.
