Conway: Claude's Always-On Agent OS Emerges
Anthropic's Conway creates persistent Claude agent environments with webhooks, extensions, and browser integration; paired with no-flicker Claude Code, GLM-5V Turbo's screen vision, and Qwen 3.6 Plus's 1M token context for production agents.
Persistent Agent Environments Unlock Always-On Workflows
Anthropic's Conway launches as a dedicated sidebar instance, not a chat session, creating a full agent workspace with chat, search, and system sections. Install custom tools and UI tabs via CNW ZIP files, building an extension ecosystem like an app store. Connectors link external clients and tools, with Chrome integration pulling browser activity into the agent loop. Webhook URLs enable event-triggered wake-ups, allowing background operation without user prompts—ideal for agentic workflows tying into Claude Code and Epitaxi controls. This native platform competes with OpenClaw by embedding deeper model integration, turning Claude into a reactive operator.
No-Flicker Mode Stabilizes Long AI Coding Sessions
Claude Code v2.1.88 introduces no-flicker mode (env var: CLAUDE_CODE_NO_FLICKER=1), using a full-screen buffer like Vim or htop to update only visible content. This eliminates flickering, stabilizes CPU/memory in extended sessions or multi-agent runs. Full mouse support adds clickable cursor positioning, expandable tool outputs, direct URL/file opens, drag-select copy, smooth scrolling, and protocol-aware behaviors (e.g., Kitty/WezTerm Ctrl+C copies). Trade-offs include adjusted native search (Cmd+F), but it bridges terminal friction for graphical-like DX, preferred by internal users.
Screen-Aware Vision Models Handle Messy Real Inputs
Z.ai's GLM-5V Turbo combines vision and coding via CogVLM Vision Encoder for layout preservation and multi-token prediction for speed on 200k-token contexts. Trained across 30+ tasks (STEM, visual grounding, video, tools), it processes screenshots, UIs, PDFs, videos directly—bypassing text descriptions—for agent chains in OpenClaw/Claude Code. Show a bug screenshot or mockup; it suggests fixes. Benchmarks SOTA on CCBench, V-bench, ClawBench, ClawEval for multimodal coding/execution, enabling natural workflows like 'fix this screen part.'
1M-Token Context Enables Repo-Scale Engineering
Alibaba's Qwen 3.6 Plus uses hybrid architecture for efficient perceive-reason-act loops across full codebases, not snippets. Default 1M-token window tracks massive projects, instructions, histories—free preview on OpenRouter. Handles multimodal inputs (docs, screenshots, wireframes, videos) into frontend code; integrates OpenClaw, Claude Code, Klein, Wukong for enterprise agents. Boosts reliability/scalability over 3.5 series, targeting production deployment over demos.