Dynamic Workflows and Ultracode NEW

Delegate 100 agents sequentially and your context collapses first. Dynamic workflows solve this by keeping the plan in script variables rather than in Claude’s context — intermediate results stay in the script, and only the final result returns to the session. It is where tokenomics meets loop engineering: enabling massive fan-out while containing context cost.

Info

One-line summary: Dynamic workflows are automation scripts written in JavaScript that orchestrate dozens to hundreds of agents in parallel. Ultracode is triggered by /effort ultracode or the ultracode keyword.

Platform basics

Background on the platform layer is in Dynamic Workflows. This page is the MoAI-ADK account of it.

MoAI-ADK provides 3 orchestration primitives, and the selection criterion is “who holds the plan.”

MoAI’s default mode — delegating one agent per turn, in sequence.

When to use:

• Simple 1-5 agent tasks
• Coding-centric run-phase work
• When agents have many inter-dependencies

A mode where multiple teammates collaborate via a shared TaskList.

When to use:

• Multiple teammates working in parallel
• Cross-layer dependencies (backend ↔ frontend)
• Collaboration and review between teammates needed

Warning

In v3.0, MoAI’s Agent Teams static orchestration layer was retired. Forcing --team falls back to sub-agent mode. The native Claude Code teammate runtime (e.g. the GLM panes of moai cg) continues to operate.

Automation scripts written in JavaScript orchestrate many agents.

When to use:

• Large-scale parallel work (dozens to hundreds of agents)
• Whole-codebase scans
• Large migrations
• Cross-source verification

A flowchart for deciding which primitive to choose.

flowchart TD
    START[Assess task characteristics] --> Q1{How many independent
agents needed?}
    
    Q1 -->|1-5| Q2{Parallel execution
required?}
    Q1 -->|5-10| Q3{Very
complex?}
    Q1 -->|10+| WORKFLOW["Choose Dynamic Workflow
Optimal for parallel scripts"]
    
    Q2 -->|No| SUBAGENT["Sequential Sub-agent
Sequential delegation"]
    Q2 -->|Yes| TEAMS["Agent Teams
Team collaboration"]
    
    Q3 -->|Yes| TEAMS
    Q3 -->|No| SUBAGENT
    
    SUBAGENT --> DONE["✓ Selection complete"]
    TEAMS --> DONE
    WORKFLOW --> DONE

/effort ultracode

Enables automatic workflow generation for all substantive work in the current session.

Effects:

• Reasoning effort: set to xhigh
• Automatic workflow generation enabled
• The optimal orchestration primitive is chosen per task

When to use:

• Very complex multi-phase work
• Large projects that need automatic orchestration

If you want to trigger a workflow for a single request rather than the whole session, use the keyword.

> Find and classify all TODO comments in our codebase.
> (Without the ultracode keyword, runs as a regular sub-agent)

VS

> ultracode: Find and classify all TODO comments in our codebase.
> (Automatically generates a workflow)

// Workflow script: classify TODOs across the entire codebase
const packages = [
  "internal/auth",
  "internal/api",
  "internal/db",
  "pkg/utils"
];

const results = [];

for (const pkg of packages) {
  // Create an independent agent for each package
  const result = await agent({
    agentType: "Explore",
    model: "sonnet",
    effort: "low",
    prompt: `
      Find and classify all TODO comments in the ${pkg} package.
      Format: [file] [line] [category] [content]
    `
  });
  results.push({ pkg, todos: result });
}

// Final consolidation
const summary = {
  total_packages: packages.length,
  package_summaries: results,
  grand_total_todos: results.reduce((sum, r) => sum + r.todos.length, 0)
};

return summary;

Workflow agents cannot interact with the user directly.

✗ A workflow agent raising a question to the user → not possible
✓ The MoAI orchestrator collects all choices up front → then runs the workflow

Resolution:

1. The MoAI orchestrator calls AskUserQuestion
2. Collects the user’s responses
3. Runs the workflow with the responses included in its input

Workflow execution requires user approval just like any run phase. A massive fan-out does not make the human gate disappear.

/moai run --workflow SPEC-XXX

→ MoAI: "Running this SPEC as a workflow. Proceed?"
→ AskUserQuestion approval required

Dynamic workflows save context, but total token consumption can be large. The fan-out scale is the cost.

Cost controls:

• Model: use sonnet low effort (read-only extraction)
• Agent count: limit scope (packages.slice(0, 20))
• Parallelism: manually tune down from the max of 16

Dynamic workflows run only under the following conditions.

1. Claude Code v2.1.154+
2. A paid plan (Pro or Team)
3. "disableWorkflows": false in /config

Can be disabled at the organization or user level.

/config
# Turn off the dynamic workflows toggle

OR

export CLAUDE_CODE_DISABLE_WORKFLOWS=1

• Builder Agents and Harness v4 - dynamic team creation
• Agent Guide - agent system overview
• SPEC-Based Development - integrated workflow

Info

Tip: For small workloads, Sequential Sub-agents suffice. Use dynamic workflows only when you need to “orchestrate dozens to hundreds of independent tasks in parallel” — and remember that the fan-out itself is the cost.