Key Signals

• Claude Code v2.1.71 introduces scheduled prompt execution, moving the tool toward always-on development assistance. The new /loop command runs any prompt or slash command on a recurring interval (e.g., /loop 5m check the deploy), while companion cron scheduling tools enable persistent automation within a session. Combined with dozens of memory leak fixes and startup performance improvements, this release signals Anthropic's push to make Claude Code viable for long-running, unattended workflows — a prerequisite for true agentic development. [1]

• Copilot SDK achieves backward compatibility across protocol versions, easing adoption for ecosystem builders. Version 0.1.32 allows SDK applications written against the v3 API to seamlessly connect to v2 CLI servers with zero code changes. Paired with v0.1.31's strongly-typed PermissionRequestResultKind constants for .NET and Go, these releases lower the barrier for tool makers integrating with the Copilot CLI ecosystem while maintaining type safety across language boundaries. [2][3]

• Open-source coding agents are challenging proprietary tools by decoupling the agent layer from the model provider. OpenCode's new $10/month Go tier — powered by cost-effective models from Chinese AI labs (Zhipu's GLM-5, Moonshot's Kimi K2.5, MiniMax M2.5) — demonstrates that LLM economics have fallen enough to support low-margin agent subscriptions. With OpenCode at 117K GitHub stars and Cline at 58.7K, the open-source agent layer is emerging as a meaningful alternative to vendor-locked tools like Codex and Claude Code. [4]

• Google's new Workspace CLI is purpose-built for AI agents, signaling a shift toward "Agent DX" as a design paradigm. The gws tool dynamically generates its command surface from Google's Discovery Service, outputs structured JSON, supports MCP, and includes 100+ pre-built agent skills and explicit OpenClaw integration. Its self-updating architecture — no static command list, no version bumps — is designed for agents operating over long time horizons, and the 14,000 GitHub stars it accumulated before media coverage suggests developers already understand what it enables. [5]

• OpenClaw community momentum continues to outpace its security posture. ClawCon NYC drew 700+ attendees for the first stop on a global meetup tour, with sponsor Kilo Code reporting 7,000 sign-ups for its KiloClaw wrapper in two days. But security researcher audits found roughly 15% of OpenClaw's skill repository contains malicious instructions, and incidents like an agent deleting a Meta employee's inbox underscore the gap between enthusiasm and safety. Core maintainers are urging users not to run agents on personal machines, while NanoClaw offers a Docker-based isolation model as a more secure alternative. [6][7]

• Anthropic deepens its enterprise strategy with a zero-commission Marketplace for Claude-powered software. Launch partners Snowflake, Harvey, and Replit can sell through the platform without Anthropic taking a revenue cut — a deliberate contrast to the 3–15% commissions charged by AWS and Azure marketplaces. The launch came hours after the Pentagon formally designated Anthropic a supply-chain risk, and appears designed to consolidate enterprise lock-in by letting committed API spend flow to third-party tools without separate procurement cycles. [8]

AI Coding News

• ClawCon NYC drew over 700 attendees to celebrate the OpenClaw open-source AI assistant platform. The event at Ideal Glass Studios in Manhattan featured demos from OpenClaw "wrappers" — one-click onboarding tools that simplify agent setup. Main sponsor Kilo Code announced 7,000 sign-ups for its KiloClaw tool within two days. Attendees ranged from a Columbia PhD student automating his neuroscience lab to an engineer scraping e-commerce data from Chinese and Japanese markets. Despite the enthusiasm, presenters repeatedly emphasized security risks: core maintainer Vincent Koc devoted a slide entirely to "Security. Security. Security." and urged users never to run agents on personal machines. Creator Peter Steinberger, now at OpenAI, was absent, prompting audience speculation about OpenAI's relationship with the project. [6]

• NanoClaw isolates each AI agent in its own Docker container, offering a security-first alternative to running OpenClaw directly. Each NanoClaw agent starts with no knowledge of other agents and accesses only explicitly mapped resources, addressing the security concerns that have plagued OpenClaw since its debut. Setup involves Claude Code for configuration, Docker Desktop for containerization, and Slack API tokens for communication. The containerized architecture proves the agent is genuinely isolated — initial attempts to access host folders failed until the mappings were correctly configured — demonstrating that security and convenience can coexist, albeit with more engineering effort. [7]

• Open-source coding agents like OpenCode, Cline, and Aider are reshaping AI dev tool economics by operating as a neutral layer above LLM providers. OpenCode's new $10/month Go tier bundles access to models from Zhipu, Moonshot, and MiniMax — Chinese AI labs whose models cost significantly less to run than Western frontier systems. At 117K GitHub stars, OpenCode has become one of the fastest-growing open-source projects, up from 44.6K at end of 2025. Anthropic recently tightened Claude restrictions to prevent third-party agents from routing Claude Code subscriptions, while OpenAI models remain accessible inside tools like OpenCode — a growing competitive divide in the model provider landscape. [4]

• Google published gws, a unified CLI for all Google Workspace APIs built explicitly for AI agents and humans. The tool collapses Gmail, Drive, Calendar, Docs, Sheets, and other services into a single command-line interface that outputs structured JSON and supports Model Context Protocol. Its architecture is uniquely self-updating: commands are generated dynamically from Google's Discovery Service, so new API endpoints appear automatically. The repository includes 100+ pre-built agent skills and explicit OpenClaw integration instructions — a notable choice given that OpenClaw's architect Peter Steinberger recently joined OpenAI. Google notes it is not an officially supported product. [5]

• Anthropic launched a zero-commission enterprise Marketplace for third-party Claude-powered software. Enterprise customers with committed API spending can redirect a portion toward partner applications from Snowflake, Harvey, and Replit without additional procurement cycles or commission fees. The no-commission structure stands in stark contrast to AWS and Azure's 3–15% marketplace fees, signaling that Anthropic currently values enterprise lock-in over transaction revenue. The launch arrived 24 hours after the Pentagon formally designated Anthropic a supply-chain risk — the first time this label has been applied to a domestic US company rather than a foreign adversary. [8]

• AI is not killing open source, but it is actively undermining the assumptions open-source projects have relied on for decades. AWS Head of Open Source Strategy Stormy Peters warned at the Linux Foundation Members Summit that AI-generated code has 1.7× more issues than human-written code, and that AI makes expert developers 19% slower due to time spent reviewing and tweaking nearly-correct output. The cURL project received 16 AI-generated security bounty submissions in eight hours, none of which identified real vulnerabilities. Some projects have banned AI-generated code entirely, while others like Fedora welcome responsible use — but only 30% of contributors disclose when they use AI. Peters characterizes this as a "one-to-two-year problem" requiring urgent attention. [9]

• OpenAI's Codex has reached 1.6 million weekly active users following its Windows launch, while two new open-source agent orchestration frameworks emerged. OpenAI Symphony monitors work queues, spawns agents in isolated environments, runs tests, and submits pull requests. Paperclip takes a broader approach, managing entire organizations of agents with org charts, budgets, and spending limits. Separately, Anthropic's annualized revenue run rate reportedly hit $19 billion (up from $9 billion at end of 2025), driven by a surge after the Pentagon dispute pushed 2.5 million users toward a "QuitGPT" movement and Claude to #1 on Apple's App Store. [10]

Feature Update

• Claude Code v2.1.71 adds /loop for recurring prompt execution and cron scheduling tools for persistent automation. The new /loop command accepts a duration and a prompt (e.g., /loop 5m check the deploy), running the specified task on a recurring interval within a session. The voice system gains a rebindable voice:pushToTalk keybinding, and eleven additional Unix commands join the bash auto-approval allowlist. The release addresses over 20 bug fixes including a stdin freeze in long-running sessions, a 5–8 second startup freeze caused by CoreAudio initialization, forked conversations sharing plan files, and plugin installations being lost when running multiple instances. Bridge session reconnection after laptop wake now completes in seconds instead of up to 10 minutes. VS Code gains a spark icon for session management, a full markdown plan viewer with comment support, and a native MCP server management dialog. [1]

• Copilot SDK v0.1.32 adds backward compatibility with v2 CLI servers, enabling v3 SDK applications to connect without code changes. This closes a significant migration gap: developers building against the latest v3 API can now also serve users still running older v2 CLI versions. The companion Go module (go/v0.1.32) adds the same v2 protocol backward compatibility adapters. [2][15]

• Copilot SDK v0.1.31 introduces strongly-typed PermissionRequestResultKind for .NET and Go, completing parity across all four language SDKs. Developers can now use typed constants like PermissionRequestResultKind.Approved and PermissionRequestResultKind.DeniedInteractivelyByUser instead of comparing against magic strings. Additional changes include get_last_session_id() / GetLastSessionID() for Python and Go, a new timeout parameter for Python RPC methods to override the default 30-second timeout for long-running operations, and a fix for Go's PermissionRequest fields which were previously a generic Extra map[string]any catch-all. [3][16]

• Gemini CLI v0.34.0-nightly.20260307 ships a unified KeychainService, a generic CacheService, and standardized MCP tool naming. The nightly build migrates token storage to a unified KeychainService, implements a generic CacheService to optimize setupUser, and refactors all MCP tool names to a standardized mcp_ fully-qualified name format. Plans are now preserved during chat compression, screen reader support is added to the RewindViewer, and the --all flag is implemented for extensions uninstall. The v0.33.0-preview.5 stable release cherry-picks a critical MCP tool naming fix. [11][12]

• OpenCode v1.2.21 fixes line ending preservation in the edit tool and resolves Git path resolution on Windows. The release addresses a PTY session handle leak, adds an interactive timeline visualization feature, and includes improvements to the desktop app such as sidebar reveal animations, hover peek overlays, and a fix for file tree tab persistence. Twelve community contributors participated in this release, including fixes for /export toggling in the TUI, Safari autocorrect handling, and session header guards. [13]

• OpenAI Codex ships four Rust alpha releases (0.112.0-alpha.8 through .11) on March 7. The releases, published over the course of the day, continue the rapid iteration cadence for the Rust-based Codex rewrite. No detailed release notes were provided for these alpha builds, consistent with the project's current pre-release development approach. [14]