4 Updates from Anthropic, OpenAI, and Google: The Hidden Pattern

The developer community built the patterns first. Then the products swallowed them whole.

Three companies shipped coding agent updates in close succession this week, and the directions overlap in ways that aren’t coincidental. Here’s what changed and the thread connecting all of it.

Claude Code Remote Control Replaces the tmux + Tailscale Combo

Terminal command /remote-control hands your active session to your phone. This sounds small until you consider what developers were actually doing before it existed: spinning up a tmux session, routing through Tailscale, and hoping the connection stayed alive when the laptop screen went dark. Cumbersome enough that several open-source tools (Happy and VibeTunnel among them) emerged specifically to solve it.

Anthropic watched that ecosystem form, then built the same thing natively.

• Research preview, currently for Max plan users
• QR code scan establishes the mobile connection instantly
• Auto-reconnects after the laptop sleeps
• Local filesystem and any running MCP servers stay intact

The third point matters more than it sounds. Previous workarounds often lost the local context on reconnect. This keeps everything in place.

Cowork Enterprise Expansion: The Plugin Marketplace Is the Actual Battleground

When Cowork launched as a research preview in January, the reaction in developer circles was something like “interesting, but what does it actually do?” This update answers that.

Admins can now create internal plugin marketplaces and deploy different configurations per team. Ten role-specific plugin templates were added (HR through investment banking), and cross-app integration now lets Excel analysis flow directly into PowerPoint, skipping the manual export-import loop most enterprise users know too well.

The deeper play here:

• 12 new connectors including Docusign and FactSet
• Structured input forms for slash commands
• OpenTelemetry-based usage tracking
• Plugins use the same format as the Claude Agent SDK

That last point is the tell. Anthropic is converging the developer-facing and enterprise-facing surfaces toward a common plugin format. The marketplace is a distribution mechanism for whatever gets built on that SDK.

Codex Sub-Agents and Responses API WebSocket: Separate but Compounding

Two changes in one release, worth separating out.

Codex now officially supports sub-agents: complex tasks get split across parallel agents that work simultaneously and merge results. This has been achievable in custom setups for a while, but /experimental flag access means it’s now a first-class feature.

The Responses API WebSocket mode is the more immediately impactful change for anyone running multi-step pipelines. The old model: every tool call meant a new HTTP connection and resending the full conversation context from scratch. WebSocket keeps a single connection open with server-side in-memory state. New inputs only.

The numbers from Cline team testing:

• ~40% speed improvement on complex multi-file coding tasks
• ~15% on simpler tasks
• WebSocket connections persist up to 1 hour before needing a reconnect
• Compatible with store=false and Zero Data Retention

For tasks with 20+ tool calls, the latency reduction is significant. Each tool call no longer pays the full context retransmission cost.

These two changes compound: sub-agents running in parallel over WebSocket connections see multiplicative latency improvements compared to either change alone.

Gemini CLI Adds Hooks: Agent Loops You Can Actually Control

In v0.26.0, Google added what Claude Code shipped last September: hooks. Insert security checks before a tool executes. Force a test run at session end. The behavior is now in config, not in a prompt.

There’s a meaningful difference between writing “always run the linter” in a system prompt and enforcing it through a BeforeTool hook. One is a suggestion. The other is a guarantee.

• AfterAgent hooks enable “Ralph loop” implementations (agents that verify their own output)
• Hooks can be bundled into extensions for team distribution
• Claude Code hook scripts migrate with minor modifications
• Three-tier config priority: project, user, system

The migration note matters for anyone already running Claude Code hooks in production. The formats are close enough that the transition is low-friction.

The common thread

Every one of these updates absorbed something the community was already doing.

Remote access via tmux and Tailscale was a real workflow before Anthropic formalized it. Plugin ecosystems existed in community form before Cowork’s marketplace. Sub-agent patterns were being hand-rolled in custom pipelines before Codex added the flag. Hook systems were being manually implemented before Gemini CLI made them configurable.

None of these companies invented the behaviors. What they’re competing on is the speed at which they observe what developers are doing and turn it into product. The model is commoditizing faster than anyone predicted. Workflow absorption is the actual moat.

The real battleground in AI tools isn’t the model. It’s how quickly a platform can see what people are doing around it and pull that into the core product.