Anthropic unleashes parallel subagent swarms in Claude Code

AI coding tools have spent the last two years acting like smart autocomplete or friendly pair programmers. If you needed to migrate an entire legacy service or rewrite a massive framework, you still needed a team of human engineers and a quarterly roadmap. Anthropic is trying to change that math.

The company just launched dynamic workflows inside Claude Code, enabling the assistant to tackle massive engineering tasks end-to-end. Instead of relying on a single agent trying to juggle an entire codebase, Claude writes dynamic orchestration scripts that spin up tens to hundreds of parallel subagents in a single session.

The feature is available across the Claude Code CLI, Desktop app, VS Code extension, and the Claude API, including integrations on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. It is on by default for Max, Team, and Enterprise tiers, while Pro users can enable it manually in settings.

750k lines of Rust in eleven days

The headline showcase for this release is ambitious. Bun creator Jarred Sumner used dynamic workflows to port Bun from Zig to Rust, spanning roughly 750,000 lines of Rust code in just eleven days from first commit to merge, hitting a 99.8% pass rate on the existing test suite.

The process was broken down systematically. One workflow mapped Rust lifetimes for every struct field in the Zig codebase. Another dispatched hundreds of parallel agents to write matching Rust files, assigning two independent reviewer agents to scrutinize every single file.

A continuous fix loop kept running builds and tests until everything was green. Claude even executed an overnight cleanup workflow to strip unnecessary data copies and opened PRs for each. While Sumner noted the code is not yet running in production, the turnaround time is wild for a project of that scale.

Hundreds of agents under one hood

Running hundreds of agents usually causes chaos, context rot, and hallucinations. Anthropic handled this by pulling task orchestration entirely outside the main conversation thread.

When a workflow starts, Claude breaks the problem into discrete pieces and runs subagents in parallel. Crucially, the system uses adversarial verification: independent agents explore solutions from different angles while others actively attempt to disprove findings until the answers converge. Runs can stretch across hours or days, with saved checkpoints so interrupted jobs pick up right where they stopped.

The launch also ties directly to the new Claude Opus 4.8 model, which Anthropic claims is four times less likely than its predecessor to let code flaws slip through unnoticed.

Users can trigger workflows manually or toggle on a dedicated setting called ultracode within the effort menu. This sets the effort tier to xhigh and lets Claude autonomously decide when a task warrants a multi-agent workflow. The trade-off is resource usage: dynamic workflows chew through significantly more tokens, prompting a required confirmation screen on the first run.

Why it matters

For engineering teams, this changes the playbook for dreaded, backlog-clogging chores. Large-scale security hardening, framework deprecations, and dead-code sweeps used to require weeks of developer time. Now, a single lead engineer can orchestrate comprehensive repo refactoring without having to wire up custom agent frameworks like LangGraph from scratch.

Whether you should dive in immediately depends on budget. Spinning up dozens of subagents on xhigh effort will burn tokens at an eye-watering pace. We think smaller teams should start strictly with scoped, low-risk modules before handing over larger enterprise repositories.

For the broader AI ecosystem, this marks a clear transition from chat-based assistants to autonomous software engineering teams. By baking parallel orchestration, adversarial checking, and persistent state directly into the CLI, Anthropic is setting a brutal bar for developer tooling competitors.

Anthropic made a smart architectural bet by isolating multi-agent coordination outside conversational memory. But there is an obvious caveat: passing 99.8% of tests does not equal production stability, as Bun's team acknowledged. If a codebase lacks bulletproof test suites to act as guardrails, dynamic workflows could easily generate subtly broken code at massive scale.