Claude Code Dynamic Workflows: Hundreds of Agents, One Session

What Dynamic Workflows Actually Does

Traditional Claude Code sessions are powerful but serial. One agent, one context window, one task at a time. For routine coding work that's fine. For a codebase-wide security audit, a 500-file framework migration, or a language port at scale, serialization becomes the bottleneck.

Dynamic Workflows removes that bottleneck by moving the orchestration plan from Claude's context into a JavaScript script that a background runtime executes. Claude writes the script based on your task description. The runtime fans work across subagents in parallel. Intermediate results live in script variables — not Claude's context — so the plan stays coherent no matter how large the task grows.

The Verification Loop: What Makes It Different

Parallel execution is table stakes in 2026. What separates Dynamic Workflows from simply running multiple Claude Code sessions is the adversarial verification layer.

Here's how a workflow run progresses:

1. Plan phase — Claude reads your task description, decomposes it into independent units of work, and writes an orchestration script
2. Fan-out phase — The runtime spawns parallel subagents, each handling an independent unit
3. Adversarial review — A second set of agents independently challenges the primary agents' outputs, looking for errors, edge cases, and incorrect assumptions
4. Convergence — The run continues iterating until results from primary and adversarial agents agree
5. Integration — Verified results are folded together; you receive a single coordinated answer

Real-World Use Cases and When to Use Them

Anthropic is explicit about where Dynamic Workflows shines. The pattern is consistent: many independent units of work verifiable by existing tests.

Strong fits

Large-scale migrations — Framework upgrades, API deprecations, language ports spanning thousands of files. Claude handles decomposition, parallel execution, and test-gate verification.

Codebase-wide audits — Security reviews (OWASP), dead code scans, missing test coverage analysis. Independent agents cover separate domains while adversarial agents challenge each finding.

Critical refactors requiring verification — When a single overlooked file means a production bug, the adversarial loop reduces that risk structurally rather than relying on human review alone.

Stress-testing plans — Draft an architecture proposal from multiple independent angles before committing. The workflow surfaces contradictions you'd find weeks later in implementation.

Weak fits

Tasks with tight cross-file dependencies that can't be parallelized, codebases with weak or absent test coverage (the quality gate disappears), and work requiring a single coherent architectural decision across the entire codebase.

Comparing Orchestration Options in Claude Code

Dynamic Workflows is one of three multi-agent patterns now available in Claude Code. Choosing the right one depends on what you know upfront.

The decision rule: if you can name the roles upfront, Agent Teams fit. If you don't yet know how to decompose the task, Dynamic Workflows fit.

How to Get Started

Activation: Dynamic Workflows are on by default for Max and Team plan users. Enterprise users need admin activation in Claude Code settings.

Starting a workflow: Two methods.

# Method 1: Ask directly in your Claude Code session
"Create a workflow to audit this codebase for OWASP vulnerabilities"

# Method 2: Enable ultracode mode for automatic workflow planning
/effort ultracode

Ultracode mode sets effort to xhigh and lets Claude decide when a task warrants a workflow. It applies to every task in the session — token usage increases accordingly.

Claude Opus 4.8: The Model Underneath

Dynamic Workflows launched alongside Claude Opus 4.8, which brings specific improvements relevant to unattended runs. Anthropic reports Opus 4.8 scores 0% on uncritically reporting flawed results and is 4x less likely than 4.7 to let a defect pass unflagged. For a workflow running hundreds of agents with no human in the loop, that honesty improvement translates directly into fewer false passes reaching you.

Pricing is unchanged from Opus 4.7: $5/M input, $25/M output. Fast Mode runs at 2.5× speed and is now 3× cheaper than the previous generation's Fast Mode: $10/M input, $50/M output.

The Broader Shift

Dynamic Workflows represents a meaningful change to what "a single Claude Code session" can accomplish. Work that previously required sprint-level planning — a full codebase security audit, a major framework migration, a large-scale refactor — now has a different answer: describe the goal, turn on a workflow, review the output.

The combination of automatic decomposition, parallel execution, adversarial verification, and checkpoint-based resumability removes the four blockers that made unattended large-scale engineering automation impractical. Whether this proves transformative in production depends on test suite quality — the verification loop is only as reliable as the tests it runs against.