knows commands. Agent: makes decisions.

Claude Code = Agent.

https://t.co/njT8U12fhj

• [twitter] "@AbdMuizAdeyemo: wait...so Claude Enterprise seat cost doesn't include usage limit? so I pay for seats and then pay separately for usage??..." (2026-04-16T06:09:44.000Z) URL: https://x.com/AbdMuizAdeyemo/status/2044659488684974186 wait...so Claude Enterprise seat cost doesn't include usage limit? so I pay for seats and then pay separately for usage??

This is going to make people freak out when they see their bills." (2026-04-16T06:09:44.000Z)

• [twitter] "@OKAMOTO_LINE: 今日初めてClaude Codeでアプリ開発してみた。最初はプロンプトで最終形を思い描いてたけど、これってバイブコーディングとは似て非なるものなのかなって思った。Claude Codeは言語でアプリを作るもので、バイブとはニュアンスが違うのかなって。..." (2026-04-16T06:09:38.000Z) URL: https://x.com/OKAMOTO_LINE/status/2044659465040044034 今日初めてClaude Codeでアプリ開発してみた。最初はプロンプトで最終形を思い描いてたけど、これってバイブコーディングとは似て非なるものなのかなって。Claude Codeは言語でアプリを作るもので、バイブとはニュアンスが違うのかなって。。バイブコーディングはもっと感覚的なものかなって。
• [reddit] "Claude Code with UltraPlan just took over my project planning workflow" (2026-04-16T04:53:28.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smu0su/claude_code_with_ultraplan_just_took_over_my/ I've been meaning to write this up for a while. Last month I switched my entire project planning workflow to Claude Code + UltraPlan, and the shift has been pretty dramatic.

The short version: I us

• [reddit] "Claude Code Managed Agents — questions [megathread]" (2026-04-16T05:27:07.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smu8fq/claude_code_managed_agents_questions_megathread/ Megathread for questions about Claude Code Managed Agents.

A Managed Agent is a Claude Code agent that Anthropic hosts and manages for you. When you connect a repo to a Managed Agent, it handles infrastructure automatically—no server setup, no containers to maintain.

A Managed Agent is configured with:

• Instructions (CLAUDE.md)
• Set of MCP servers
• Set of allowed tools

Trigger options:

• GitHub events (push, PR open, issue label, etc.)
• API calls
• Schedule (via Routines)

Managed Agents are available to Claude Max (5x and 20x) and Enterprise plan subscribers.

More info: https://docs.anthropic.com/en/docs/claude-code/managed-agents

• [reddit] "Claude Code's new /usage command is a game changer for Max subscribers" (2026-04-16T04:41:43.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smtrqn/claude_codes_new_usage_command_is_a_game_changer/ The new /usage command in Claude Code is actually really useful. Before this update, the only way you could see your usage was the vague little bar at the top of the terminal. Now when I type /usage i

  • Shows remaining capacity as a percentage (updated in real time)
  • Breakdown by token type (input, cache read, cache write, output)
  • What's contributing to your limits (top prompts, tools, agents)
  • Estimated time until limits reset
  • Session history with cost per session

• [reddit] "Claude Code Routines: the full breakdown (trigger types, config options, use cases)" (2026-04-16T04:00:27.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smso6y/claude_code_routines_the_full_breakdown_trigger/ A quick breakdown of Claude Code Routines for anyone who hasn't dug into them yet.

  What are Routines? Routines let you schedule and automate Claude Code agents to run on a recurring basis—without you being at your computer. They run in the cloud (via Managed Agents), so your local machine doesn't need to be on.

  Trigger options:

  • Cron schedule (e.g. every weekday at 9am)
  • GitHub events (push, PR opened, issue labeled, etc.)
  • API call (trigger manually or from another service)

  What they can do:

  • Run any agent task (write code, open PRs, post comments, update docs)
  • Use MCP servers and allowed tools
  • Report results back to Slack, GitHub, or other services via MCP

  Availability:

  • Managed Agents (and thus Routines) are available on Claude Max 5x, Max 20x, and Enterprise plans

  Config:

  • Routines are configured via the Claude Code desktop app or API
  • Instructions set via CLAUDE.md file in repo

  Real-world use cases floating around:

  • Nightly dependency checks
  • Automated PR reviews triggered by push
  • Weekly digest generation (e.g. IR summaries, changelog drafts)
  • Monitoring pipelines: detect anomaly → comment on issue → open fix PR

• [reddit] "Claude Code Desktop: The "Mission Control" redesign makes parallel sessions actually usable" (2026-04-16T05:03:24.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smua5g/claude_code_desktop_the_mission_control_redesign/ Big quality of life improvement in the Claude Code Desktop redesign.

  The new "Mission Control" layout puts all your active Claude Code sessions in one persistent sidebar. Each session has its own:

  • Worktree (separate git branch + working directory)
  • Model selection
  • Tool permissions
  • Status indicator (running, waiting, done, error)

  From Mission Control you can:

  • Launch new sessions in parallel
  • Switch between them without losing context
  • Kill or restart individual sessions
  • See which session is using which MCP server

  Previously, parallel sessions were technically possible but required juggling multiple terminal tabs and mentally tracking which was doing what. The Mission Control sidebar makes it genuinely practical to run 3-4 parallel agents on different features simultaneously.

  Worktrees are now first-class in the UI—each session automatically gets its own branch so merges are clean.

• [reddit] "Claude Code v2.1.92 changelog notes" (2026-04-16T03:17:14.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1sms6gt/claude_code_v2192_changelog_notes/

  Major additions in v2.1.92:

  UltraPlan

  • New planning mode that generates and displays a full project plan before beginning execution
  • Plan is editable by user before Claude commits to it
  • Runs the planning phase in cloud (Managed Agent infrastructure) even if local machine is offline
  • Aiming to reduce wasted compute on wrong-direction work

  Managed Agents (GA)

  • Agents that Anthropic hosts and runs for you — no infra required
  • Configurable via CLAUDE.md + MCP servers + allowed tools
  • Trigger via: GitHub events, API, or scheduled Routines

  Routines

  • Schedule agents on cron, GitHub event triggers, or API calls
  • Cloud-run, no local machine needed
  • Available to Claude Max (5x, 20x) and Enterprise plan subscribers

  Stream stability fix

  • Streams now have a 5-minute keep-alive timeout (previously would silently drop on long-running tasks)

  Default model change

  • Claude Code now defaults to recommending Sonnet 4.6 instead of Opus 4.6 1M

  UI/UX

  • /usage command now shows token breakdown + usage contributors
  • "Always allow" button repositioned in permission dialogs
  • Desktop app: Mission Control sidebar for parallel session management

  Breaking / migration notes:

  • If you rely on the default model recommendation being Opus, you will need to explicitly set your preferred model
  • Existing stream-dependent integrations may need testing against the new 5-min keepalive behavior

• [reddit] "Opus 4.7 spotted on Google Vertex" (2026-04-16T03:10:39.000Z) URL: https://i.redd.it/t93hibcrygvg1.png

• [reddit] "Claude Code's UltraPlan is actually Ultraplan" (2026-04-16T03:41:39.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smsbgh/claude_codes_ultraplan_is_actually_ultraplan/ Wait, I've been calling it the wrong thing this whole time. It's "Ultraplan" not "UltraPlan." Lowercase p.

  Anyway it's great. Made a mistake on day 1 by diving into complex implementation without making a plan first. Claude started heading in the wrong direction and I had to restart. After switching to the Ultraplan workflow, I now spend ~10 min reviewing the plan, make a few tweaks, then let it go. Way fewer mid-session corrections needed.

• [reddit] "Claude Enterprise: seat fees do NOT include usage — billed separately at API rates" (2026-04-16T04:09:46.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smsrwq/claude_enterprise_seat_fees_do_not_include_usage/ Flagging this for anyone evaluating Claude Enterprise for their team.

  Seat fees cover platform access only. All usage — prompts, responses, Claude Code sessions, Cowork interactions — is billed separately at standard API rates.

  There is no per-seat usage allowance. You pay for exactly what your team consumes.

  Source: Anthropic's Claude Enterprise docs (confirmed by Grok)

• [reddit] "Agentic coding report: 60% of dev work uses AI, but only 23% let agents run autonomously — anyone else find that gap surprising?" (2026-04-16T05:57:25.000Z) URL: https://reddit.com/r/ClaudeAI/comments/1smv0qw/agentic_coding_report_60_of_dev_work_uses_ai_but/ Pulled from Anthropic's 18-page agentic coding report.

  Key numbers:

  • 60% of developer work now involves AI assistance
  • Only 23% of that involves fully autonomous agent runs
  • Most developers are still in "copilot mode" — they prompt, review, prompt again

  The gap surprised me. With tools like Claude Code's Managed Agents and Routines now available, I'd expect more autonomous usage. But the data suggests the bottleneck is trust, not tooling.

  Two theories:

  1. Devs don't trust agents to not break things (especially in prod-adjacent contexts)
  2. The tooling for review/rollback isn't mature enough yet to feel safe

  Anyone here actually running autonomous agents in prod? How do you handle oversight?

Claude Code Changelog Articles:

• [changelog] "Claude Code v2.1.92 released" (2026-04-16) Summary: UltraPlan (editable pre-execution plans), Managed Agents GA, Routines (scheduled agents), stream stability fix (5-min keepalive), default model changed to Sonnet 4.6, /usage command improved, Mission Control desktop redesign.

AI Daily Briefing — April 16, 2026

Today's digest is dense with Claude Code momentum — Anthropic shipped v2.1.92 and the community is buzzing across four continents. Meanwhile, researchers are rethinking RL fundamentals and the LLM political bias benchmarking space is heating up.

Claude Code Developer Corner

v2.1.92 — The Agentic Developer's Release

Claude Code v2.1.92 dropped and it's one of the more consequential updates in recent memory. Multiple Reddit threads, a community changelog breakdown, and a wave of international Twitter chatter confirm this is landing widely. Here's what's new:

Ultraplan (editable pre-execution planning) The headline feature. Before Claude starts executing on a complex task, it now generates a full project plan and surfaces it for your review — cloud-side, even if your machine is offline. You can edit the plan, redirect scope, or approve it before any code is touched. Early adopters report dramatically fewer mid-session corrections. One community note: it's officially spelled "Ultraplan" (lowercase p), not "UltraPlan."

Managed Agents (GA) Anthropic-hosted agents that handle all infrastructure for you. Connect a repo, configure via CLAUDE.md, set allowed tools and MCP servers, pick a trigger — done. No containers, no server maintenance. Trigger options:

• GitHub events (push, PR open, issue labeled, etc.)
• API calls
• Scheduled Routines (see below) Available on Claude Max 5x, Max 20x, and Enterprise plans.

Routines (scheduled agents) Routines let you automate recurring agent tasks — nightly dependency checks, auto-PR reviews on push, weekly IR summary drafts, anomaly detection pipelines that open fix PRs. Cloud-run, no local machine required. Configured via the desktop app or API.

Mission Control Desktop Redesign The Claude Code desktop app now has a "Mission Control" sidebar showing all active sessions in parallel. Each session gets its own worktree (separate git branch + directory), model selection, tool permissions, and status indicator. You can launch, switch between, kill, or restart sessions without losing context. Worktrees are now first-class UI citizens — previously managing parallel sessions meant juggling terminal tabs.

Stream Stability Fix Long-running tasks previously dropped streams silently. Now: 5-minute keep-alive timeout. If you have stream-dependent integrations, test against the new behavior.

Improved /usage Command No longer just a vague progress bar. /usage now shows:

• Remaining capacity as a percentage (real-time)
• Token breakdown by type (input, cache read/write, output)
• Top contributors to your limits (prompts, tools, agents)
• Estimated time until reset
• Session history with per-session cost

Breaking Changes / Migration Notes

• Default model shift: Claude Code now recommends Sonnet 4.6 by default, not Opus 4.6 1M. If you rely on Opus by default, explicitly set your preferred model. (Community has noticed this — see @fakedev9999.)
• Stream behavior: test any integrations that rely on long-running streams against the new 5-min keepalive.
• Permission dialog: the "Always allow" button has moved — relearn muscle memory.

Enterprise Pricing Clarification One piece of important context for teams evaluating Claude Enterprise: seat fees do NOT include usage. All Claude Code sessions, prompts, and Cowork interactions bill separately at standard API rates. Plan accordingly.

Token Efficiency Note A community analysis circulating on Twitter puts Claude Code at 5.5x better token efficiency vs. Cursor for equivalent tasks — 82% fewer tokens, translating to roughly $400/mo vs. $2,200/mo at scale. Take with appropriate skepticism, but the direction of the gap tracks with other reports.

What You Can Do Now That You Couldn't Before

• Run agents while your laptop is closed (Routines + Managed Agents)
• Edit Claude's plan before it touches your code (Ultraplan)
• Manage 3-4 parallel agent sessions in one UI (Mission Control)
• See exactly what's eating your token budget (new /usage)
• Auto-trigger agent workflows on GitHub push/PR/issue (Managed Agents)

Research & Benchmarks

LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning — A new benchmark targeting LLMs deployed on complex autonomous tasks, focusing on planning accuracy over extended reasoning chains. As agent use cases grow, long-horizon CoT reliability becomes a first-order concern.

From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space — Researchers argue that RLVR (RL with verifiable rewards) is fundamentally bounded by the base model's unconditional distribution P(y). The implication: pushing RLVR harder without improving pretraining may be hitting a ceiling.

From Feelings to Metrics: Understanding and Formalizing How Users Vibe-Test LLMs — A formalization of "vibe-testing" — the informal, experience-based eval approach most practitioners actually use. The paper attempts to bridge the gap between benchmark scores and real-world usefulness perception.

LLM Politics & Model Behavior

Political Compass Benchmark for Frontier LLMs — A community-built benchmark using 98 structured questions across 14 political dimensions maps frontier LLMs on a 2D political compass. Notable finding: Kimi K2 refuses all questions about Taiwan, and GPT-5.3 refuses 100% of questions when given an opt-out. Methodology is worth scrutinizing, but the results surface real policy-behavior divergences worth tracking.

AI in Education

Why Sal Khan's AI Revolution Hasn't Happened Yet — Sal Khan reflects on Khanmigo's rollout and why the expected classroom AI transformation has been slower than predicted. Friction points: teacher adoption, institutional inertia, and the gap between demo quality and daily-use reliability. Worth reading for anyone building AI tools targeting education.

Agentic AI Trends

Anthropic's Agentic Coding Report: 60% AI Involvement, Only 23% Autonomous — Anthropic's 18-page agentic coding report finds that ~60% of developer work now involves AI, but only 23% involves fully autonomous agent runs. The gap suggests the bottleneck is trust and review tooling, not capability. With Managed Agents and Routines now GA, it will be interesting to see if that 23% moves.

Worth Watching

• Opus 4.7 spotted on Google Vertex — Unconfirmed, but a screenshot surfaced showing Opus 4.7 listed in the Vertex model catalog. No announcement yet from Anthropic.
• Claude Enterprise billing surprise — Seat fees don't include usage; everything bills at API rates. Finance teams evaluating enterprise deployment need to model this carefully.
• GraphQL MCP Server — Open-source tool that auto-discovers your GraphQL schema and generates MCP tools with zero config. Supports stdio and HTTP modes.
• PPO multi-timescale advantage routing causes policy collapse — Undergrad researcher documents a failure mode when dynamically routing multi-timescale advantages in PPO, and proposes a decoupled fix. Niche but practically useful for RL practitioners.
• TREX: Agent-Driven LLM Fine-Tuning via Tree-Based Exploration — Automates the LLM fine-tuning workflow using an agent that explores a tree of training strategies. Early but interesting for teams trying to automate their training pipelines.
• Claude-Brief: Session digest tool for Routines — A community-built tool that generates a digest of what Claude Code agents did overnight: files changed, commits made, failures, and cost. Useful complement to Routines.

Sources

• Claude Code v2.1.92 changelog — (internal changelog)
• Claude Code Routines: the full breakdown — https://reddit.com/r/ClaudeAI/comments/1smso6y/claude_code_routines_the_full_breakdown_trigger/
• Claude Code Managed Agents megathread — https://reddit.com/r/ClaudeAI/comments/1smu8fq/claude_code_managed_agents_questions_megathread/
• Claude Code Desktop: Mission Control redesign — https://reddit.com/r/ClaudeAI/comments/1smua5g/claude_code_desktop_the_mission_control_redesign/
• Claude Code v2.1.92 changelog notes — https://reddit.com/r/ClaudeAI/comments/1sms6gt/claude_code_v2192_changelog_notes/
• Claude Code's UltraPlan is actually Ultraplan — https://reddit.com/r/ClaudeAI/comments/1smsbgh/claude_codes_ultraplan_is_actually_ultraplan/
• Claude Code with UltraPlan just took over my project planning workflow — https://reddit.com/r/ClaudeAI/comments/1smu0su/claude_code_with_ultraplan_just_took_over_my/
• Claude Code's new /usage command is a game changer — https://reddit.com/r/ClaudeAI/comments/1smtrqn/claude_codes_new_usage_command_is_a_game_changer/
• Claude Enterprise: seat fees do NOT include usage — https://reddit.com/r/ClaudeAI/comments/1smsrwq/claude_enterprise_seat_fees_do_not_include_usage/
• Agentic coding report: 60% of dev work uses AI — https://reddit.com/r/ClaudeAI/comments/1smv0qw/agentic_coding_report_60_of_dev_work_uses_ai_but/
• Anthropic's agentic coding report (Reddit summary) — https://i.redd.it/nmj774tylhvg1.jpeg
• @fakedev9999 on default model change to Sonnet 4.6 — https://x.com/fakedev9999/status/2044663201816785366
• @BuzzRag on Claude Code v2.1.92 — https://x.com/BuzzRag/status/2044662170659815702
• @Oxymarun_ on Anthropic's last-month shipments — https://x.com/Oxymarun_/status/2044661739464302771
• @musiol_martin on token efficiency (Claude Code vs Cursor) — https://x.com/musiol_martin/status/2044661315441242466
• @jxkedevs on claude-brief — https://x.com/jxkedevs/status/2044663718126960882
• @nifeio on GraphQL MCP Server — https://x.com/nifeio/status/2044661166128177197
• @rawattsumit on Claude Code Routines — https://x.com/rawattsumit/status/2044661006614610268
• @ins_aisolutions on Claude Code Desktop redesign — https://x.com/ins_aisolutions/status/2044662325253132513
• @hobbydevelop on /usage update — https://x.com/hobbydevelop/status/2044661400858173641
• @aiagent_builder on Ultraplan/Monitor — https://x.com/aiagent_builder/status/2044662964935106815
• Opus 4.7 spotted on Google Vertex — https://i.redd.it/t93hibcrygvg1.png
• LongCoT: Benchmarking Long-Horizon CoT Reasoning — http://arxiv.org/abs/2604.14140v1
• From P(y|x) to P(y): RL in Pre-train Space — http://arxiv.org/abs/2604.14142v1
• From Feelings to Metrics: Vibe-Testing LLMs — http://arxiv.org/abs/2604.14137v1
• Political Compass Benchmark for LLMs — https://reddit.com/r/MachineLearning/comments/1smqsbu/built_an_political_benchmark_for_llms_kimi_k2/
• Why Sal Khan's AI Revolution Hasn't Happened Yet — https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/
• PPO multi-timescale advantage routing failure — https://reddit.com/r/MachineLearning/comments/1smr52p/why_dynamically_routing_multitimescale_advantages/
• TREX: Agent-Driven LLM Fine-Tuning — http://arxiv.org/abs/2604.14116v1
• Claude Enterprise billing (Grok confirmation) — https://x.com/grok/status/2044660542460023125