Anthropic's New Claude CONWAY Is Unlike Any AI Before

Key Concepts

• Agentic Workflows: AI systems designed to perceive, reason, and execute multi-step tasks autonomously.
• Persistent Environments: AI workspaces that maintain state and run in the background rather than operating as transient chat sessions.
• Multimodal Coding: The ability of AI to process visual inputs (screenshots, UI layouts, videos) and translate them directly into functional code.
• Context Window: The amount of data (tokens) a model can process and "remember" at one time; critical for repository-level engineering.
• No-Flicker Rendering: A terminal UI technique using full-screen buffers to stabilize performance and improve user experience in AI coding tools.

---

1. Anthropic: Conway and Claude Code Updates

Anthropic is shifting Claude from a conversational model to a persistent platform.

• Conway (The "Always-On" Agent):
  • Architecture: Operates as a dedicated environment/OS for agents rather than a standard chat. It features a sidebar interface with sections for Search, Chat, and System management.
  • Extensibility: Introduces CNW ZIP files, an ecosystem for custom tools and UI tabs, allowing developers to package functionality similar to an app store.
  • Integration: Includes a web-hook system that allows external services to trigger the agent, enabling it to run in the background. It also features a "connected client" system, allowing browser-based Claude to interact directly with the Conway instance.
  • Control Layer: References to "Epitaxi" suggest a centralized control layer for managing these persistent agents.
• Claude Code "No-Flicker" Mode (v2.1.88):
  • Technical Fix: Replaces standard terminal rendering with a full-screen buffer (similar to vim or htop), eliminating screen jumping and reducing CPU/memory overhead.
  • UX Improvements: Adds full mouse support (click-to-position, drag-to-select, scroll-wheel navigation) and advanced clipboard integration, making the terminal feel like a graphical interface.

2. Z.AI: GLM 5V Turbo

Z.AI focuses on bridging the gap between visual perception and actionable coding.

• Core Capability: Unlike models that convert images to text descriptions before reasoning, GLM 5V Turbo processes visual input (UI layouts, bug screenshots, wireframes) directly.
• Technical Architecture:
  • Cogvit Vision Encoder: Preserves fine visual detail and layout structure.
  • MTP (Multi-Token Prediction): Enhances speed and output length management.
• Real-World Application: Optimized for OpenClaw and Claude Code, allowing developers to point at a visual bug or mockup and receive code suggestions based on the visual context.
• Benchmarks: Claims state-of-the-art performance on multimodal coding benchmarks like CCbench V2, Zclaw Bench, and Claw Eval.

3. Alibaba: Qwen 3.6 Plus

Alibaba’s latest release targets enterprise-grade, large-scale agentic deployment.

• Repository-Level Engineering: Designed to handle entire codebases rather than isolated snippets, facilitating complex, multi-file tasks.
• 1 Million Token Context: A massive context window that allows the model to maintain long-term memory of project files, tool usage, and instruction chains.
• Deployment: Integrated into Alibaba’s Wukong enterprise platform. It is also compatible with existing agent frameworks like OpenClaw, Claude Code, and Klein.
• Hybrid Architecture: Focuses on efficiency and scalability, with improved reasoning reliability compared to the 3.5 series.

---

Synthesis and Conclusion

The industry is undergoing a fundamental transition from "AI as a Chatbot" to "AI as an Operator."

• Anthropic is building the infrastructure for persistent, background-running agents that act as their own operating systems.
• Z.AI is solving the "visual reasoning" bottleneck, allowing AI to interact with the messy, visual reality of software development.
• Alibaba is providing the "memory" and scale required for enterprise-level, repository-wide automation.

The common thread across these developments is the move toward agentic autonomy: systems that can perceive a problem (via screen or code), reason through the steps, and execute the solution without constant human intervention. The focus has shifted from model intelligence to workflow integration and system stability.