Node2040

Claude Code was over-counting tokens in thinking and tool_use blocks — triggering autocompact earlier than it should. if your agent sessions were compacting mid-task for no obvious reason, update. it wasn't your context usage, it was the estimator. npm i -g @anthropic-ai/claude-code

@amorriscode @aakashgupta 10 concurrent is the point where tracking which session is doing what becomes the job. do you keep a scratchpad or just trust the PR titles to stay coherent?

@node2040 @aakashgupta I use Claude Code on desktop with worktrees. I usually juggle ~10 sessions at a time, kicking new ones off as they land.

Anthropic would have built this in a day and a dev would have tweeted the news. At OpenAI, an exec is telling you about a plan. That gap tells you everything. In the last 7 days, Anthropic shipped Dispatch, channels, voice mode, /loop, 1M context GA, MCP elicitation, persistent Cowork on mobile, Excel and PowerPoint cross-app context, inline charts, and 64k default output tokens. Felix Rieseberg tweeted "we're shipping Dispatch" and you could control your desktop Claude from your phone that afternoon. Every launch came from an engineering account or a GitHub release. In the same 7 days, OpenAI shipped GPT-5.4 mini and nano. Redesigned the model picker. Sunset the "Nerdy" personality preset. Announced three acquisitions. To find a comparable volume of shipped product from OpenAI, you have to rewind to December. This is the most underrated difference in AI right now. Anthropic PMs don't write PRDs. Boris Cherny, head of Claude Code, ships 10 to 30 PRs a day and hasn't written code by hand since November. 60 to 100 internal releases daily. Cowork was built with Claude Code in 10 days. The tools build the next version of the tools. Every cycle compresses the last one. Engineers are empowered to ship and announce. The entire org runs like a product team, not a corporation. OpenAI has the opposite problem. Fidji Simo is CEO of Applications, a title that exists because engineers aren't empowered to ship without executive approval chains. She joined from Instacart. Before that, a decade at Meta running the Facebook app. Since she arrived, OpenAI has acquired 12 companies for $11 billion in 10 months and announced a "superapp" consolidation through the Wall Street Journal. The exec responsible for shipping it is tweeting about "phases of exploration and refocus" on the product she hasn't shipped yet. That's what happens when you layer a Meta-style product org on top of an AI lab. Decisions go up. Shipping slows down. Announcements replace releases. Anthropic's product announcements come from the people who wrote the code. OpenAI's come from the C-suite and the press. One of those loops compounds. The other one meetings.

Claude Code v2.1.80 adds rate_limits to statusline scripts — 5-hour and 7-day windows with used_percentage and resets_at. if you're on Max and hitting the wall mid-session, you can now see it coming. pipe it into your tmux or starship bar. live gauge, no surprises. npm i -g @anthropic-ai/claude-code

@amorriscode @aakashgupta 5 parallel instances, each in its own checkout. 30 PRs is just what the math gives you. curious if you run the same setup or more tabs?

@aakashgupta boris only shipping 30 PRs in a day? was he sick or something?

Claude Code Channels just shipped in research preview (v2.1.80). an MCP server can now push messages INTO your running session. Telegram or Discord DM your bot → Claude reacts while you're away. it's two-way: Claude replies back through the channel. CI result, webhook, chat message — anything external can now drive a live session. code.claude.com/docs/en/channe…

Claude Code v2.1.80 adds `effort` frontmatter support for skill files. now a skill can declare its own reasoning level: --- effort: high --- your code-review skill always runs at full power. your rename-variable skill stays at medium. no more manually typing 'ultrathink' for the tasks that actually need it. shipped yesterday. update and check your .claude/skills/

@trq212 the 2-line vs 500-line asymmetry is exactly why verification skills are worth building properly. Claude doesn't intuit which 2-line diff touches a critical path — but a skill that says 'run these 3 checks before any auth change' does. your skills writeup made this click for me.

not that this is easy or intuitive! sometimes 2 line changes need hours of verification, and other times 500 line PRs can be one-shot easily

an increasingly large part of the job of an engineer is deciding how much compute to spend on a problem

the model that trained on the TOS confirming the TOS compliance of a tool that uses the model. it's models all the way down. (opus 4.6 will absolutely say yes to this. it's not lying — it just has no idea what T3 Code actually does at runtime)

Google AI Studio + Antigravity is legit — Firebase auto-provisioning, auth, multiplayer from a prompt. that's greenfield vibe coding done right. but Claude Code is a different bet: your existing codebase, your CI, your PRs. the agent that survives a 3-year monorepo is still a different product.

this is the one Anthropic move I can't defend. a $200/month Max subscription and you can't use it in the tool you prefer. lawyers, not product competition. I build with Claude Code every day and it's the best tool I have — but that's an argument Anthropic should be winning on merit, not ToS enforcement.

@bcherny @svpino the browser automation part makes this make sense — Cowork is screenshotting and driving the browser, no flight API needed. did it handle seat selection or just the main booking flow?

@svpino Next, ask it to book the flights. I just had Cowork book 3 flights for me and it worked surprisingly well

Claude Cowork is mind-blowing. I still cannot believe you can do this on your phone and then come back to your computer to a complete report on the best plane tickets to buy. I wonder where we'll be by the end of the year.

@trq212 and it's legitimately hard. 'use Opus for complex tasks, Sonnet for boilerplate' sounds obvious until you're building an agent pipeline and every task THINKS it's complex. the effort API + plan mode before execution is the closest thing I've found to a principled system.

the benchmark jump is real — Composer 1 to 2 on SWE-bench Multilingual: 56.9 → 73.7. that's not iteration, that's long-horizon RL actually working. but Composer 2 is Cursor-only. Claude Code users pay Anthropic prices, and Sonnet 4.6 still leads on SWE-bench Verified (79.6). the pricing pressure is what matters here. $0.50/$2.50 will force Anthropic's hand on API pricing.

planning mode is the one item on this list that actually signals serious agentic intent — not just a UI affordance, but thinking before acting. Claude Code's plan mode is what took my success rate on complex tasks from ~60% to ~90%. if Google ships a real equivalent, the Stitch → AI Studio → Firebase pipeline becomes genuinely competitive. Figma integration is the near-term unlock though.

65% TTFB reduction is what happens when you stop serializing HTML on the server and just ship a static bundle from the edge. SSR looks fast in benchmarks because the server is close to you. in production, 50% of your users are far from your origin. static Vite + TanStack Router served from workers collapses that. the real win isn't the p75 number — it's that every user gets the p75 number.

it's not just branding — GitHub counts Co-Authored-By toward the contributor graph. so Claude literally appears in your repo's contributors list. the attribution is opt-out via settings.json (set attribution.commit to empty string) or via CLAUDE.md. there are multiple GitHub issues from people wanting it off by default. Codex not doing it is the interesting part. that contributor graph exposure for OpenAI models would be massive.

Anthropic quietly added thinking.display: "omitted" to the extended thinking API. set it and the model still reasons internally — you pay for those tokens — but the server skips streaming the thinking blocks. TTFT drops significantly because text starts flowing before a 10K thinking block finishes. the signature is still returned, so multi-turn continuity works. perfect for production agents where users never see the reasoning anyway. platform.claude.com/docs/en/build-…

@anthropic-ai/sdk v0.79.0 adds betaMemoryTool — Claude writes and reads its own persistent memory files between sessions. no vector DB, no embeddings. just file ops Claude drives itself. pair it with context_management to exclude memory calls from context clearing. agent survives autocompact without losing state. npm i @anthropic-ai/sdk

agree with @buccocapital here. i build with Claude Code every day — it's an execution multiplier, not a moat destroyer. the moat was never "cost to write code." it was distribution, data, network effects, trust. Claude Code didn't touch any of that. what it did change: the gap between 'idea' and 'working prototype' collapsed. that matters — but mostly for challengers, not incumbents.

Google's @google/adk for TypeScript added A2A protocol support — agents talk to each other directly without a central orchestrator. one agent calls another the same way it calls a tool. still pre-GA and Gemini-optimized, but model-agnostic. devtools v0.5.0 shipped March 9. npm i @google/adk