AI Coding News: July 22, 2026 — Claude Code Caps Subagents at 20, Codex CLI Stabilizes Multi-Agent Mode
Four releases landed across the major AI coding assistants in the last two days. Claude Code's v2.1.217 puts a default ceiling on how many subagents can run at once and closes a budget-enforcement gap. GitHub shipped two back-to-back Copilot CLI point releases fixing an infinite-loop bug and a command-approval leak between repositories. OpenAI's Codex CLI graduated its multi-agent mode out of preview and added audio input. And OpenCode's v1.18.4 brings adaptive thinking controls to Kimi models while continuing its Desktop v2 polish.
Claude Code v2.1.217: Subagent Concurrency Gets a Default Cap
Claude Code's fifth release in six days, v2.1.217, lands July 21 and continues the week's focus on taming multi-agent workloads. The headline change is a default cap of 20 concurrently-running subagents, overridable with the CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS environment variable — a hard ceiling on top of the subagent nesting-depth limits Claude Code shipped earlier this month. A related fix closes a gap where --max-budget-usd wasn't stopping background subagents once the budget ran out, meaning a runaway agent fleet could previously burn past a configured spend limit. Elsewhere in the release: a memory leak where truncated MCP tool outputs were retaining their full, untruncated results in memory is fixed; Windows auto-update failures that could leave claude.exe missing from disk now restore the previous executable instead; corporate mTLS, TLS-verify, OAuth scope, and proxy settings that were being silently ignored are respected again; and background session isolation now canonicalizes symlinked working directories correctly. On the lighter side, the prompt input gained emoji shortcode autocomplete — type :heart: for ❤️, toggle off via emojiCompletionEnabled.
GitHub Copilot CLI v1.0.72 and v1.0.73: Closing an Agent-Stop Loop and a Cross-Repo Leak
GitHub shipped two Copilot CLI point releases back to back. v1.0.72 (July 20) fixes an infinite-loop bug in the agentStop hook that could leave a blocking agent stop spinning indefinitely, and stops command approvals from leaking between repositories — previously, approving a command in one repo's session could carry over into another repo, a meaningful isolation gap for anyone running Copilot CLI across multiple projects. v1.0.73 (July 21) follows up with two fixes for custom agents: Anthropic-style subagents now keep working correctly when additional directories are configured, and relative links inside custom agent instruction files now resolve from the agent file's own location instead of breaking.
OpenAI Codex CLI rust-v0.145.0: Multi-Agent Goes Stable, Audio Arrives
OpenAI's Codex CLI rust-v0.145.0 (July 21) is the biggest release of the day. Its multi-agent V2 experience — configurable sub-agent models, reasoning levels, concurrency, and roles — graduates from opt-in preview to stabilized, alongside audio input and tool-output support for common local audio formats and new streaming realtime V3 conversations. The release also adds experimental paginated thread history with efficient resume, search, persisted names, and memories; expands /import to pull in settings, MCP servers, plugins, sessions, commands, and project-scoped memories from both Cursor and Claude Code, useful for anyone consolidating onto Codex from another tool; and adds experimental Amazon Bedrock login and custom-endpoint support with GPT-5.6 Sol as the default Bedrock model. On the reliability side, editing an earlier prompt or retrying a safety-buffered turn now creates a proper contextual branch instead of mutating the original conversation, and Windows sandboxing gets native exec-server support and correctly quoted hook commands.
OpenCode v1.18.4: Adaptive Thinking for Kimi, Azure Endpoint Restored
OpenCode's v1.18.4 (July 20) adds adaptive thinking controls for Kimi models running on Anthropic-compatible providers, with summarized reasoning output on by default. It also restores Azure Cognitive Services endpoint support that had regressed, and reduces OpenAI provider header timeouts that were slowing down connection setup on slow networks. On the desktop side — which finished its v2 layout migration just days earlier in v1.18.0 — this release syncs the embedded terminal theme with the app theme, keeps open file tabs aligned with the current diff in the review panel, and rewrites the v2 prompt input for more reliable command and context handling.
Why It Matters
The through-line across today's four releases is the one that's been running through this whole cycle: as agentic coding tools push further into multi-agent and long-running background work, every vendor is spending real engineering effort on the guardrails around it rather than on flashier headline features. Claude Code capped subagent concurrency and closed a budget-enforcement gap; Copilot CLI fixed a cross-repo approval leak that matters more as usage scales across projects; Codex stabilized its multi-agent mode only after enough hardening to trust it outside preview; and OpenCode kept polishing its new desktop UI while extending model support. If you run Claude Code with background subagents, update to v2.1.217 and check whether the new default cap of 20 concurrent subagents fits your workflows. If you use Copilot CLI across multiple repositories in the same session, update to at least v1.0.72 to close the approval-leak gap. And if you've been waiting for Codex's multi-agent mode to leave preview before relying on it, rust-v0.145.0 is that signal.
Sources & Further Reading
Comments
Share your thoughts and join the conversation
Leave a Comment
Keep reading.
Daily Dev Stack Release Audit — July 28, 2026: Redis, MongoDB, and Nuxt Rush Out Critical Security Patches
AI Coding News: July 28, 2026 — Model Context Protocol Ships Its Final 2026-07-28 Spec
