AI Coding News: August 2, 2026 — Claude Code Defaults to Opus 5, Copilot CLI Locks Down Sandboxes

Three of the most-used AI coding agents shipped meaningful updates this week, and the common thread is agent safety: default models got smarter, sandboxes got stricter, and background-agent plumbing got more reliable. Here is what actually changed, sourced directly from each project's changelog.

Claude Code 2.1.219: Opus 5 Becomes the Default Model

Anthropic's Claude Code CHANGELOG shows version 2.1.219 introducing Claude Opus 5 as the default Opus model, now with a 1M-token context window — a significant jump for teams working across large monorepos. The release also ships several changes aimed squarely at agent safety and orchestration:

• A new sandbox.network.strictAllowlist setting for tighter network access control during agent runs.
• A DirectoryAdded hook that lets a mid-session working directory register without restarting the agent.
• Nested subagent forwarding enabled in stream-json output, with spawning now supported up to depth 3 by default.
• A new workflowSizeGuideline settings key for dynamic multi-agent workflow sizing, plus mcp_server_errors added to headless stream-json init events for better MCP diagnostics.

The follow-up build, 2.1.220, is a smaller reliability and bug-fix release. Full details are in the official "What's new" docs.

GitHub Copilot CLI Locks Down Sandbox Defaults, Adds Grok 4.5

GitHub's Copilot CLI changelog shows two back-to-back releases tightening how the agent handles permissions. Version 1.0.77 (July 30) makes unconditional autopilot approval disable the sandbox for the current session when bypass is allowed, switches interactive terminals to a web-based OAuth login flow by default, enforces sandbox-denied paths for relative and symlinked entries, and adds support for the grok-4.5 model. The prior release, 1.0.76 (July 29), added enable/disable controls in /plugins for plugins, instructions, agents, LSP servers, and hooks, an experimental sessions sidebar for managing concurrent sessions, and the ability for enterprise administrators to enforce restrictive sandbox policies.

Separately, GitHub announced a default model enablement policy for Copilot Business and Enterprise: newly GA models will now be turned on automatically instead of requiring admins to manually enable each one. Organizations that want to keep the old opt-in behavior need to disable the policy before August 26.

OpenCode v1.18.11 Fixes MCP Reconnect Loops

The open-source, terminal-native agent OpenCode shipped v1.18.11 on August 1, per its GitHub releases page. The release resolves MCP SSE connections that were stuck in reconnect loops after server errors, corrects provider model configurations using interleaved reasoning fields, makes external links open in the system browser instead of inside the app, and fixes desktop file-tree tab clipping during resizing. The prior build, v1.18.10 (July 30), added automatic discovery of available Modal models and improved toast notification stacking on mobile.

Why This Matters

Sandboxing and network allowlists used to be an afterthought in coding agents. Now three of the most widely used tools — Claude Code, GitHub Copilot CLI, and OpenCode — are all shipping stricter default permission handling in the same week. For teams running these agents with broad autonomy, it's worth reviewing the new allowlist and sandbox-policy settings before your next upgrade rather than after.

Sources & References

• Claude Code CHANGELOG (anthropics/claude-code)
• Claude Code "What's new" docs
• GitHub Copilot CLI changelog
• GitHub Changelog: Default model enablement for Copilot Business and Enterprise
• OpenCode releases