Key Signals

• GitHub opens Claude and Codex coding agents to all Copilot tiers, completing a multi-agent platform play. Copilot Business and Pro subscribers can now run Claude and OpenAI Codex alongside GitHub's own Copilot agent — directly inside github.com, GitHub Mobile, and VS Code — with no extra subscription cost. Each agent session consumes one premium request and operates on a shared platform with unified governance, shared context, and shared memory. This effectively turns GitHub into a multi-vendor agent marketplace where teams can assign competing agents to the same issue and compare outputs. [1]

• GitHub's Enterprise AI Controls and agent control plane reach general availability, establishing the governance layer for agentic development at scale. The release introduces searchable agentic session activity filtered by specific agents, API support for enterprise-wide custom agent definitions, and a permanent AI Controls tab in enterprise settings. Combined with the same-day expansion of Claude and Codex access, this signals GitHub's strategy: democratize agent access while giving enterprises the audit logging, MCP allowlists, and policy controls they need to adopt agents safely. [2]

• Andrej Karpathy proposes replacing "vibe coding" with "agentic engineering," reflecting a material shift in how AI agents are used in production. Industry analysts from Forrester, Futurum Group, and Constellation Research agree that coding agents have undergone a "huge step change" in the last 1-2 months, now capable of building fully working compilers via multi-agent parallelism. The consensus: the differentiator is no longer the model but the "agentic harness" — the system of tools, context management, evaluation, and observability surrounding the agent. Cognitive debt is emerging as a new concept alongside technical debt. [3]

• Cursor ships Bugbot Autofix, achieving a 35% merge rate on automated PR fixes — a concrete benchmark for autonomous code repair. Bugbot now goes beyond identifying issues in pull requests to running cloud agents on dedicated machines that test changes and propose fixes directly. Users can configure autofix to either post preview comments or push changes directly to the branch. This represents a quantifiable step toward fully autonomous code review and repair loops. [4]

• A wave of CLI tool releases lands across Codex, Copilot CLI, Claude Code, Gemini CLI, and OpenCode, collectively advancing memory, security, and multi-agent orchestration. Codex 0.106.0 adds a direct install script, diff-based memory forgetting, and GPT-5.3-codex visibility; Claude Code v2.1.59 introduces auto-memory and a /copy command; Copilot CLI v0.0.419-1 adds mouse-mode controls and MCP naming improvements; Gemini CLI's nightly adds an experimental browser agent and runtime hooks. The convergence of memory management, MCP interoperability, and sub-agent support across all major tools indicates rapid platform maturation. [5][6][7][8]

• Microsoft opens two fronts in the agentic AI battle: Copilot Tasks for consumer automation and the Agent Framework RC for developer orchestration. Copilot Tasks uses its own cloud-based computer and browser to handle recurring, scheduled, or one-time busywork described in natural language — Microsoft's direct response to Claude Cowork, ChatGPT Agent Mode, and Perplexity Computer. Meanwhile, the Microsoft Agent Framework hits Release Candidate for .NET and Python, consolidating Semantic Kernel and AutoGen into a single production-ready SDK with MCP interoperability, graph-based multi-agent workflows, and streaming support. [9][10]

• OpenAI Codex and Figma launch a code-to-design integration, bridging the gap between implementation and visual iteration. The new partnership enables teams to move between code and the Figma canvas for faster iteration and shipping. This extends Codex's reach beyond pure code generation into the design-development workflow — a historically siloed boundary that agentic tools are beginning to dissolve. [11]

AI Coding News

• Andrej Karpathy suggests retiring "vibe coding" in favor of "agentic engineering," sparking industry debate about how to describe AI-assisted development. A year after Karpathy popularized the term, experts say the practice has outgrown the label. Forrester analyst Diego Lo Giudice notes this was predicted in Q4 2024: "Vibe coding will transform into vibe engineering by the end of 2026." Arcjet CEO David Mytton observes that agents now complete complex tasks autonomously, but only when guardrails — documentation, tests, structured workflows — are in place. Multiple voices argue that 2026's primary engineering challenge is "cognitive debt," not just technical debt. [3]

• Microsoft previews Copilot Tasks, a cloud-based AI agent system that operates its own computer and browser to handle background busywork. Users describe tasks in natural language and Copilot Tasks executes them on a recurring, scheduled, or one-time basis — from organizing subscriptions and drafting email replies to planning events and monitoring apartment listings. The system asks permission before meaningful actions like payments. Currently in research preview with a waitlist, Copilot Tasks is Microsoft's answer to the wave of agentic products from Anthropic, OpenAI, Perplexity, and Google. [9]

• Microsoft Agent Framework reaches Release Candidate for .NET and Python, unifying Semantic Kernel and AutoGen into a single production-ready SDK. The framework enables developers to create AI agents in just a few lines of code, with function tools for type-safe domain calls, graph-based workflows for multi-agent orchestration, and MCP interoperability. Built-in support for streaming updates, checkpointing, and human-in-the-loop scenarios positions this as a serious enterprise alternative. Packages remain pre-release on NuGet and PyPI while migration guides from Semantic Kernel and AutoGen are being finalized. [10]

• OpenAI and Figma launch a Codex integration connecting code and design workflows. The partnership enables teams to move between code implementation and the Figma canvas for faster iteration, addressing the traditional friction between design and development. This marks another expansion of Codex's footprint beyond terminal-based coding into visual design tooling. [11]

Feature Update

• GitHub Copilot now offers Claude and Codex as coding agents for all Copilot Business and Pro users. Previously limited to Enterprise and Pro+ tiers, the expansion makes Claude and OpenAI Codex available across github.com, GitHub Mobile, and VS Code at no additional cost. Each agent session consumes one premium request. All three agents — Copilot, Claude, and Codex — share a unified platform with shared governance, context, memory, and audit logging. Users can assign issues to multiple agents simultaneously, mention @copilot, @claude, or @codex in PR comments, and compare agent approaches side by side. [1]

• GitHub Enterprise AI Controls and the agent control plane are now generally available. New capabilities include searchable agentic session activity across all agents, API support for enterprise-wide custom agent definitions via .github-private/agents/*.md, and the permanent AI Controls tab in enterprise settings replacing the old Copilot policies page. Cloud agent session activity now exceeds the prior 1,000-record limit. MCP enterprise allowlists remain in public preview as GitHub designs cross-organization scaling. [2]

• GitHub Copilot Content Exclusion REST API enters public preview. Organization and enterprise administrators can now programmatically manage content exclusion rules — specifying files, directories, or repositories that Copilot should not reference when generating suggestions. The JSON API supports GET operations for listing and managing exclusion patterns, enabling automation of sensitive-content governance at scale. [12]

• GitHub Mobile launches Live Coding Agent Notifications for real-time agent session tracking. Using Live Activities on iOS 17.2+ and Android 16+, developers see session states for Copilot, Codex, and Claude agent sessions. Users can tap through to the pull request for immediate review. Only cloud-run sessions are supported; the feature can be disabled via Settings or a one-tap opt-out on the first notification. [13]

• GitHub Copilot metrics API migrates download URLs to a new endpoint. Report data, API contract, and response schema are unchanged. Organizations consuming Copilot usage metrics programmatically should verify integrations still work, particularly if consuming URLs directly rather than following redirects. [14]

• Cursor ships Bugbot Autofix, enabling automated PR issue fixing with a 35% merge rate. Autofix runs cloud agents on dedicated machines to test changes and propose fixes directly on pull requests. Users can configure it to post preview comments with a @cursor merge command or push changes directly to the branch with no interaction required. The feature is available from the Bugbot dashboard. [4]

• Copilot CLI v0.0.419-1 adds keyboard navigation, mouse-mode controls, and MCP improvements. New additions include Ctrl+F/Ctrl+B for page scrolling, a --mouse/--no-mouse flag, Home/End for alt-screen buffer navigation, and Ctrl+G for external editor prompts. The AUTO theme now uses the terminal's actual ANSI color palette. MCP server env vars in command/args/cwd fields are auto-included, server names now support dots, slashes, and @ characters, and /mcp enable works for previously auto-disabled built-in servers. [5]

• Claude Code v2.1.59 introduces auto-memory and an interactive /copy command. Claude now automatically saves useful context to auto-memory, and /copy shows an interactive picker for selecting individual code blocks or full responses. "Always allow" prefix suggestions now compute smarter per-subcommand prefixes for compound bash commands. Memory usage in multi-agent sessions is improved by releasing completed subagent task state. v2.1.61 fixes concurrent writes corrupting config files on Windows. [6]

• OpenAI Codex CLI 0.106.0 adds a direct install script, thread realtime API, and memory improvements. The release includes a macOS/Linux install script published as a GitHub release asset, expanded app-server v2 thread API with realtime endpoints, js_repl promoted to experimental, request_user_input enabled in Default mode, and GPT-5.3-codex visibility in the model list. Memory behavior is improved with diff-based forgetting and usage-aware selection. A security fix closes a zsh-fork sandbox bypass. [7]

• OpenCode v1.2.15 fixes Windows stability and splits TUI from server configuration. The release fixes most segfaults on Windows using Bun v1.3.10 stable, resolves a macOS hang caused by interactive shell flags in sidecar spawn, and fixes permissions handling from child sessions. Community contributions include keyboard navigation improvements and corrected Copilot provider descriptions in i18n files. [15]

• Gemini CLI nightly (v0.30.0-nightly.20260226) adds an experimental browser agent, runtime hooks, and MCP progress bars. Key additions include the experimental browser agent for Gemini CLI, RuntimeHook function support, progress bar/throttling/input validation for MCP tool progress, automatic model switching for Plan Mode, Conseca security framework, AfterTool tail tool calls, and expanded MCP tool detail display on approval. The nightly also enforces folder trust for local extension installs and adds Gemini 3.1 Pro Preview to behavioral evals. [8]