Google Antigravity Dynamic Workflows Let Agents Run as Python Scripts — Plus Claude Code 2.1.325
Google's Antigravity shipped Dynamic Workflows on 13 November, a new way to coordinate teams of AI coding agents through a Python script the developer can read before it runs. Type /run-workflow, and instead of a long conversation that drifts, the agent writes a small program with named phases and parallel subagents — then shows you the script for approval. On the same day, Anthropic released Claude Code 2.1.325, which fixes a security issue in how zstd-compressed responses are decoded and makes sandboxed Bash commands ask permission by default.
Dynamic Workflows is Google's answer to a familiar failure of multi-agent coding. Ask one agent to review fifty changed files and it works through them in a single conversation, losing track and dragging every detail into one context window until the work degrades. Antigravity's new approach hands coordination to code. The main agent loads Antigravity's built-in workflow skill and generates a Python script for the requested task. Antigravity displays the script and waits for approval before executing it. A separate Python process then runs the workflow in the background while the original conversation stays available, and when execution finishes, the main agent reads the logs and summarises the results.
How a workflow actually runs
The generated program uses a small asynchronous API built around five primitives. phase(name) starts a named phase and groups its logs and agent cards. agent(prompt, schema, role, model, workspace) defines one worker agent; awaiting it produces a single result, and keeping its handle allows follow-up turns with the same agent. parallel(items, fn) runs a function concurrently for every item and preserves result order — if one operation fails, Antigravity cancels the rest. pipeline(items, fn) processes items one after another. And await trajectory(...) makes selected earlier steps from the parent conversation available to the script.
Isolation is the design's core idea. Each workflow-launched agent receives its own conversation and cannot see the orchestration script or its peers, so the script must supply every agent with its context, its task and its output format. A subagent submits its result through a finish tool, which validates the value against a specified JSON Schema before passing the parsed result to the script. File reads, searches and test output stay inside the subagent conversations, which preserves the main conversation's context window during high-volume jobs. The official demo audits one module per worker in parallel.
![]() |
| AI-generated illustration (generated 14 Nov 2026) |
Five primitives, per-agent budgets, and Git worktrees
Developers can match cost and capability to each step. High-volume checks can use the lighter flash_lite or flash models, while tasks needing deeper reasoning can use pro. A workflow may also select Antigravity's read-only research agent or a custom subagent. That choice matters because a high-fanout workflow can trigger many model calls, and model selection is the main control over credit use.
Parallel editing gets its own safety mechanism. Assigning workspace="branch" to each agent gives every worker an isolated Git worktree, so agents can modify files and run tests concurrently without overwriting one another's changes.
Dynamic Workflows are available with Google AI Pro and Google AI Ultra, in both the Google Antigravity App and the Antigravity CLI. They have no separate price and consume the plan's existing model credits. The workflow runs in a separate process using the Python runtime bundled with Antigravity, so no local Python installation is required; developers who need a different interpreter can set the ANTIGRAVITY_PYTHON environment variable. Saving the generated program as a .py file in the workspace makes the same orchestration available for later runs, so a team can review the script alongside its code and reuse it.
Workflows, swarms, and Orchestrator are three different tools
Antigravity now offers three multi-agent options, and the task's coordination pattern decides which one fits. If the job is repeated checks or ordered stages — the same review across every changed file, a bug queue processed in order — /run-workflow is the tool, because a Python script controls execution and collects schema-validated results. If agents must negotiate or review one another's work, /swarm is the tool: agents coordinate through a shared message board. If the job is one self-contained delegated task, a single subagent handles it with less orchestration overhead.
Dynamic Workflows is also distinct from Orchestrator, which Antigravity rolled out the day before, on 12 November. Orchestrator is a model-dropdown option for Google AI Ultra subscribers that switches between Gemini 4 Carbon and Gemini 4 Hydrogen inside one session — Carbon for investigation and planning, Hydrogen for implementation and tests. Google says it reached near-Carbon task success on Terminal-Bench 4.1 while averaging 24% lower quota use and 19% faster completion. Workflows change how agents are coordinated; Orchestrator changes which model does the work inside one session.
To mark the week of shipping, Antigravity lead Varun Mohan said the team is resetting everyone's Gemini quota on Antigravity. Neither official post says how long the reset lasts, so developers should treat it as a short window.
![]() |
| AI-generated illustration (generated 14 Nov 2026) |
Claude Code 2.1.325: the zstd fix and the permission change
Claude Code is Anthropic's command-line coding agent, and version 2.1.325 arrived on 13 November with two items that matter. The first is a security fix: the release notes say, in a single sentence, "Fixed a security issue in how zstd-compressed responses are decoded." Zstd is a compression format used to shrink data in transit, and the code that expands compressed responses had a security issue. That is the whole of what Anthropic has said publicly. The notes name no CVE, list no affected versions, and describe no exploit path or impact — so the honest description stops there. Update, and do not invent a vulnerability the changelog does not describe.
The second item is a default change that many developers will feel immediately. Sandboxed Bash commands now follow the normal permission flow by default, instead of running without a prompt. Users who want the old behaviour can set sandbox.autoAllowBashIfSandboxed to true; organisations that turn the sandbox on through managed settings must set it in their highest-ranking managed source. The release also fixes related permission gaps in compound Bash commands, allow rules and relative deny rules, plus a fix for /security-review and git commands failing in worktrees of bare repositories.
The permission change is a deliberate trade: a little more friction in exchange for a clearer record of what ran. For solo developers the extra prompts are a minor annoyance; for teams, the default means sandboxed commands can no longer slip past review unnoticed. If the prompts become noisy, the documented flag restores the old behaviour — but that choice should be made deliberately, not by accident.
Background: Claude Duet is Anthropic's proactive agent for small businesses, a different product from the Claude Code command-line tool. Developers working inside Claude Code can also run a full SEO audit with the claude-seo Claude Code SEO plugin, and the Impeccable GitHub skill is the highest-starred skill in the same ecosystem. On the Google side, Gemini edit files shows how Gemini works directly inside documents and spreadsheets.
![]() |
| AI-generated illustration (generated 14 Nov 2026) |
Key Takeaways
- Google Antigravity's Dynamic Workflows, shipped 13 Nov, turn a /run-workflow request into a reviewable Python script that coordinates parallel and sequential subagents with JSON Schema-validated results.
- Five primitives define the flow — phase, agent, parallel, pipeline and trajectory — with per-agent model choice and isolated Git worktrees; the feature is on Google AI Pro and Ultra, with no separate price.
- Workflows are not /swarm (agents negotiating on a shared board) and not Orchestrator (switching Carbon and Hydrogen inside one session); each fits a different coordination pattern, and a Gemini quota reset was announced with no stated duration.
- Claude Code 2.1.325, also 13 Nov, fixes a security issue in zstd-compressed response decoding — no CVE, affected versions or exploit path were named — and makes sandboxed Bash commands ask permission by default.
- The old silent behaviour can be restored with sandbox.autoAllowBashIfSandboxed set to true, but that should be a deliberate choice, not an accident.
Conclusion
Dynamic Workflows changes the unit of work: from a conversation that drifts to a script that can be read, saved, reviewed and run again — worth more than another model switch for teams doing repeated reviews and refactors. Claude Code 2.1.325 is the smaller half of the day — a one-line security fix and a permission default that adds friction on purpose — but it points the same way: agents are getting more powerful, and the tools around them are adding the controls that let developers stay in charge. Try the workflow on a real audit this week, while the quota reset lasts, and read the generated script before you approve it.
Sources
- AlphaSignal — Google Antigravity Ships Dynamic Workflows to Run Parallel AI Coding Agents (13 Nov 2026)
- Claude Code changelog — v2.1.325 (13 Nov 2026; zstd security fix, sandboxed Bash permission default)
- Subagentic — Claude Code 2.1.325 fixes zstd decoding and makes sandboxed Bash ask first (13 Nov 2026)
- Unite.AI — Google Antigravity Rolls Out Orchestrator for Dynamic Model Handoffs (12 Nov 2026)



Comments
Post a Comment