Coding agents
Week of October 12: coding agents go team-mode
OpenCode, Codex CLI and Claude Code shipped multi-agent updates between Oct 11 and 16, after VS Code, while Meta unveiled a huge coding model.
Between October 11 and 16, 2026, three major tools shipped updates pointing the same way: running more agents, for longer, with less babysitting. VS Code had led the way a week earlier.
What shipped
- OpenCode (Oct 11-12): v2 has existed since 2.0.0, published on September 11 (npm). It introduces a background service, a new server API and a new plugin API (docs). This week, 1.19.0 on October 11 offers OpenCode 1 users the move with "opencode upgrade --major" (GitHub), and the 2.0.27 update on October 12 adds worktree tools to Code Mode (changelog).
- Codex CLI 0.163.0 (Oct 15): you can steer the agent by sending a message while it is responding, and independent read-only tools now run in parallel; the /daemon command can now restart its managed background server (GitHub).
- Claude Code 2.1.302 (Oct 16): fork subagents are on by default in claude -p, the Agent SDK, the desktop app and VS Code, as they already were in the terminal; idle compaction can be turned off from /config (GitHub).
- VS Code 1.141 (Oct 7): a session grid in the Agents window, agent sandboxing on Windows, macOS and Linux, and the ability to continue Copilot CLI or Codex conversations inside VS Code (VS Code).
On the model side
Meta launched Muse Spark Open, an 816-billion-parameter MoE model with 42 billion active per token and a 1-million-token context, built for agentic coding (NVIDIA, Oct 15). On October 19, Meta also released Muse Glimmer, a 30-billion-parameter Apache 2.0 model, quantized to about 4 bits to run on a single GPU (Meta AI). See the Muse Spark Open page.
The common thread: orchestrating, not assisting
Each vendor tackles a different pain point. OpenAI lets you correct an agent without killing the task. Anthropic makes delegating to subagents automatic. OpenCode is moving its users to a v2 built around a background service that other programs can call. Microsoft focuses on the developer watching several sessions at once.
That is the shift Gartner acknowledged when it redefined the category around agents in its first Magic Quadrant for coding (see our article). Tools no longer stand apart only on model quality, but on how they launch, isolate and control several agents.
What it changes for developers
- Read release notes before updating in CI: a changed default, such as Claude Code subagents, can alter a script without changing the command.
- Turn on sandboxing whenever an agent runs commands or reaches the network.
- Plan the OpenCode migration: v1 plugins do not run in v2, and integrations that call the old server API must be rewritten. See the OpenCode page.
- Compare agents on your own repositories: our comparison puts Codex, Claude Code and Copilot side by side.
The limits
More agents in parallel also means more logs to review, more permissions to set and more tokens to pay for. The vendors do not quantify the gains from these updates, and the Muse Spark Open figures come from a Meta partner, NVIDIA.
The trend: the single agent is giving way to coordinated agent teams.
Spotted an error in this article? Report it · Our corrections policy
