Claude Code Dynamic Workflows: Parallel Subagents & UltraCode Mode

Claude Code now supports dynamic workflows that let the agent orchestrate tens to hundreds of parallel subagents in a single session, checking work before anything reaches you. Available in research preview in the CLI, Desktop, and VS Code extension for Max, Team, and Enterprise plans, plus the API (Amazon Bedrock, Vertex AI, Microsoft Foundry).
How to use
- Ask directly: "Create a workflow"
- Enable UltraCode via the effort menu (sets effort to
xhigh), letting Claude auto-decide when to spawn a workflow.
Workflows consume substantially more tokens than a typical session — start scoped.
Capabilities shown in early access
- Codebase-wide bug hunts: parallel search across a repo, then independent verification on every finding.
- Large migrations: framework swaps, API deprecations, language ports spanning thousands of files end-to-end.
- Double-checked results: independent solution attempts and adversarial agents that try to break the result before delivering it.
Real-world example: Bun rewrite
Jarred Sumner used dynamic workflows to port Bun from Zig to Rust — ~750,000 lines of Rust in 11 days, with 99.8% of existing tests passing. Workflow steps:
- Map Rust lifetimes for every struct field.
- Write every
.rsfile via hundreds of parallel agents, each with two reviewers. - Fix loop driving build/test until clean.
- Overnight workflow to eliminate unnecessary data copies, opening a PR for each fix.
How it works
Claude plans dynamically from your prompt, breaks the task into subtasks, fans work across parallel subagents, checks results independently, and iterates until answers converge — producing a single coordinated answer.
📖 Read the full source: HN AI Agents
👀 See Also

Depct tool collects runtime data to help Claude debug production issues
Depct is a tool that collects runtime instrumentation from Node.js apps, builds graphs from the data, and feeds it to Claude via AWS Bedrock to help debug intermittent production failures. It also generates architecture diagrams and dependency maps from runtime behavior.

Self-Hosted Contextual Bandit in Rust: Syntra & Lycan for Adaptive Decision Systems
Two open-source projects: Lycan (graph execution language with strategy nodes and learned weights) and Syntra (Docker/API appliance serving compiled Lycan capsules). Found data pipeline bugs before runtime bugs when dogfooding on an AI stock-debate product.

EsoLang-Bench: A Coding Benchmark Using Esoteric Languages to Test LLM Reasoning
Researchers created EsoLang-Bench, a coding benchmark using esoteric programming languages like Brainfuck and Whitespace to test whether LLMs can reason or just pattern-match. The best result across GPT-5.2, O4-mini, Gemini, Qwen, and Kimi was 11.2%.

CC-Ledger: Track Claude Code Costs Per Session and Per PR with Local SQLite
CC-Ledger is a Rust binary that hooks into Claude Code, logging each turn to local SQLite. Catch runaway sessions live and see per-PR cost without an API key. Includes macOS menu bar, web dashboard, and CLI views.