sartoris-digital/ultrawork

Ultrawork

Ultrawork is a parallel execution engine that runs multiple agents simultaneously for independent tasks. It provides parallelism and smart model routing but is a component, not a standalone persistence mode.

When to Use

• Multiple independent tasks can run simultaneously
• User says "ultrawork", "parallel", or wants concurrent execution
• Need to delegate work to multiple agents at once

When NOT to Use

• Task requires guaranteed completion with verification — use ralph instead (ralph includes ultrawork)
• Task requires a full autonomous pipeline — use autopilot (autopilot includes ralph)
• Only one sequential task with no parallelism opportunity — delegate directly to an agent

Execution Policy

• Fire all independent agent calls simultaneously — never serialize independent work
• Route tasks to cost-appropriate tiers
• Use parallel dispatch for independent tasks, sequential for dependent ones

Model Tier Routing

| Tier | Best For | |------|----------| | fast (Haiku) | Simple lookups, definitions, trivial edits | | standard (Sonnet) | Standard implementation, feature work | | reasoning (Opus) | Complex analysis, debugging, architecture |

Steps

1. Classify tasks by independence — which tasks can run in parallel vs. which have dependencies
2. Route to correct tiers — match task complexity to model cost
3. Fire independent tasks simultaneously:

subagent({
  tasks: [
    { agent: "worker", tier: "fast", task: "Add missing type export for Config interface" },
    { agent: "worker", tier: "standard", task: "Implement the /api/users endpoint with validation" },
    { agent: "worker", tier: "standard", task: "Add integration tests for auth middleware" }
  ]
})

4. Run dependent tasks sequentially — wait for prerequisites before launching dependent work
5. Verify when all tasks complete:
   • Build/typecheck passes
   • Affected tests pass
   • No new errors introduced

Examples

Good — parallel independent tasks:

subagent({
  tasks: [
    { agent: "worker", tier: "fast", task: "Fix typo in error message" },
    { agent: "worker", tier: "standard", task: "Add pagination to GET /users" },
    { agent: "worker", tier: "standard", task: "Write tests for auth service" }
  ]
})

Bad — sequential independent tasks:

subagent({ agent: "worker", task: "Fix typo" })
// wait...
subagent({ agent: "worker", task: "Add pagination" })
// These could have run in parallel

Relationship to Other Modes

ralph (persistence wrapper)
  └── includes ultrawork (this skill)

autopilot (full lifecycle)
  └── includes ralph
      └── includes ultrawork (this skill)

Ultrawork provides parallelism. Ralph adds persistence and verification. Autopilot adds the full lifecycle.

Final Checklist

• [ ] All parallel tasks completed
• [ ] Build/typecheck passes
• [ ] Affected tests pass
• [ ] No new errors introduced