Meta opens Muse Spark, Chainguard cages agents, and Claude Code Review goes GA

1. Meta releases Muse Spark Open

Meta published weights for Muse Spark Open on October 15, calling it the most capable US open-weight model to date. It's natively multimodal, with 816B total parameters (42B active) and a 1M-token context window, aimed at long-horizon coding and agentic work. You can pull weights from Hugging Face (and cloud model gardens), or call it through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI. Meta also optimized an on-device path: two Apple M5 Ultra Mac Studios with MLX.

Our take: Try it via an API host first. The desktop story is impressive, but "open weights" still means serious hardware.

2. Chainguard Workspaces and Checks (beta)

At Assemble London, Chainguard launched two betas for agent-written code. Workspaces give each developer or agent a hardened remote microVM (chainctl develop syncs your tree, including uncommitted changes). Short-lived credentials ride a host-side proxy instead of living in the guest. Checks runs CI in warm microVMs from your lockfile, callable from the CLI, a pre-commit hook, a GitHub App, or an MCP server/SDK. Pair them and a failed check returns structured logs so the agent can retry without rebuilding the environment. Workspaces currently support Claude; other agents are planned. Both offer 30-day trials.

Our take: If your agent needs package installs and shell access, isolate it. This is the right shape of answer.

3. Claude Code Review is generally available

Anthropic moved Code Review out of research preview for Claude Team and Enterprise. Most reviews now finish in about 6–10 minutes (down from ~20 in preview). Admins pick Quick, Standard, or Deep per repository. Findings are meant to state the problem, what triggers it, and the fix in a few sentences. Eligible Enterprise orgs can claim at least $10,000 in Code Review credits through January 15, 2027 (credits expire 30 days after claim; terms apply).

Our take: Turn it on for your highest-churn repos first. Agents write faster than humans review; this is Anthropic's bid to close that gap.

4. AI-DLC from AWS Labs

AI-DLC (AI-Driven Development Life Cycle) is awslabs' open workflow engine that runs natively in Claude Code, Kiro, Codex CLI, Cursor, opencode, and GitHub Copilot. One harness-neutral core covers 5 phases and 33 stages, with human approval gates, persistent state, and a large audit trail. Install with a one-liner, then aidlc config --harness <your-tool> and aidlc doctor.

Our take: Worth it when ad-hoc agent chats keep losing the thread on anything bigger than a one-file fix.

5. T3 Code: one surface for your agent zoo

T3 Code is an open-source control plane from Ping (t3.codes / pingdotgg/t3code). It orchestrates Claude Code, Codex, Cursor, Grok Build, OpenCode, and Google Antigravity from web, desktop, and mobile, using the subscriptions you already have. Threads get their own branches, with one-button commit, push, and PR.

Our take: If you already pay for three agent CLIs, this is the remote-friendly dashboard those tools didn't ship.

Try this today

Give your agent a real definition of done. Before you kick off an overnight coding run, paste: "Done means the user flow works in the running app (or a recorded browser/API test), not that unit tests pass. Do not edit tests to make them green. Stop and report if you cannot prove the flow." Agents optimize for whatever you measure — make the measure the product, not the suite.