Claude Code's 90-Day Sprint: 35 Updates to Autonomous OS
Anthropic shipped 35 updates in 90 days, turning Claude Code from a babysat terminal tool into a hands-free OS that runs autonomously, controls desktops, and powers 4% of GitHub commits (135k daily)—via remote phone access, auto-permissions, 1M context, and managed agents at 8¢/hour.
Remote Access Transforms Workflow Independence
Claude Code now supports four ways to operate without desk presence, shipped in four weeks: (1) Remote control via QR code scan (30-second setup) lets you prompt, view actions, and approve from phone, but limited to one session with manual approvals. (2) Dispatch orchestrates multiple sessions, spinning up tools like Claude Code as needed, delivering push notifications on completion—one user achieved 3.5 hours of parallel work from 25 minutes of phone input, enabling tasks like competitive analysis over coffee. (3) Channels connect sessions to Telegram, Discord, iMessage (2-minute setup) for two-way chats and event triggers (e.g., new leads auto-start tasks). (4) Computer Use (research preview, Pro/Max plans, Mac/Windows) lets Claude view screens, move mouse, click, and automate legacy/enterprise apps via screenshots and reasoning, pairing with dispatch for phone-initiated desktop control. These fix Open Claw's edge—remote messaging—while adding screen control for un-API'd tools.
Auto-Mode and Loops Enable Trustworthy Autonomy
Permissions previously halted 93% of safe actions (file writes, installs, tests); now Auto Mode (Teams plans) uses a classifier to auto-approve them, flagging only 7% risky ones (deletes, main pushes). Workaround for Pro/Max: settings.json rules allow reads/writes/tests while blocking installs/deletes (80% effective). Loops run recurring prompts up to 3 days (e.g., /loop for CI checks); cloud scheduled tasks execute on Anthropic infra (e.g., daily mornings with Auto Mode), checkable via phone channels—no local machine needed. Result: always-on workflows without babysitting, impossible 90 days ago.
Model Upgrades and Quality Boost Reliability
Opus 4.6 (Feb 5) delivers 1M token context (full large codebases), 128k output tokens, and adaptive thinking (task-complexity-based reasoning depth), eliminating mid-task forgetting—don't max 1M tokens (performance drops), but headroom handles production code. Auto Dream consolidates memory overnight (after 24h/5 sessions), fixing contradictions, updating dates, pruning stale refs, keeping index <200 lines for efficient recall. Security review reasons like a human (traces flows, cuts false positives), uncovering 500+ long-undetected vulns in open-source repos missed by scanners. Revert to high effort level (from rushed medium) reduces hallucinations; /costs shows per-model breakdowns; MCP tool search defers loading (85-95% context savings).
Production Tools and Overlooked Power Moves
Managed Agents (public beta, Apr 8) offloads infra: API-define agent/tools/task, Anthropic sandboxes/runs (8¢/runtime hour + tokens), returns traces—ship production agents for <$1/hour without custom setup. Ultra Plan (preview) offloads 30-min cloud planning (Opus 4.6) for complex architecture, reviewable in browser before local/cloud exec. Agent teams enable parallel Cluade instances coordinating via shared lists/mailbox. Hooks inject context/pre-tool webhooks for tech stack integration. Source leak (Mar 31, 500k TypeScript lines) revealed solid architecture, earning community respect. Big picture: Claude Code (84k GitHub stars, 75% startups, doubled WAU) evolves from terminal to multi-interface OS (terminal/phone/cloud/VUI), writing 4% GitHub commits.