Luna

The False Promise of AI Orchestration AI developers are obsessed with "orchestration." The prevailing theory on tech Twitter suggests that frontier models function best as high-level managers. Instead of writing code directly, expensive models like GPT-5.6 Sol or Fable 5 act as lead engineers, delegating tasks to cheaper subagents like Luna or Sonnet 5. To test if this managerial overhead actually pays off, I designed a head-to-head experiment. I tasked these models with building a standard interactive Laravel checklist application using both a multi-agent delegation structure and a direct, single-model approach. Subagents Slow Down the Build When GPT-5.6 Sol acted as an orchestrator, it spun up multiple Luna subagents, assigning separate UI and database components to each. The workflow mirrored a real engineering team: the lead analyzed the codebase, delegated tasks, and reviewed pull requests. However, this hierarchy introduced massive latency. The entire process dragged on for 29 minutes. The main agent spent most of its time waiting for the workers, polling their progress like a driver asking "are we there yet?" and running subsequent revision cycles when subagents missed details. In contrast, running GPT-5.6 Sol as a direct implementer with high reasoning effort bypassed this communication loop entirely. The single-model build completed in just 22 minutes, running significantly faster than the multi-agent setup. Cloud Code Dominates the Interface Battle The user experience varied wildly between developer interfaces. In Codeex CLI, monitoring subagents felt like peering through a keyhole. The terminal output only showed abstract status markers like "interacted with worker." Checking individual agent progress required running manual slash commands. Cloud Code handled this workflow much better. Its interface displayed live progress panels on the side, letting me toggle between the main Fable 5 orchestrator and running Sonnet 5 workers in real time. It clearly communicated parallel operations, making the multi-agent chaos highly legible. The Shocking Math of Agent Overhead | Metric | GPT-5.6 Sol (With Agents) | GPT-5.6 Sol (Direct) | Fable 5 (With Agents) | Fable 5 (Direct) | | :--- | :--- | :--- | :--- | :--- | | **Time** | 29 mins | 22 mins | 23 mins | 13 mins | | **Resource Cost** | 24% weekly limit | 9% weekly limit | $8.84 total | $8.00 total | The real shock came from resource consumption. Delegating tasks from GPT-5.6 Sol to Luna consumed a massive 24% of my weekly Codeex CLI usage limit. Building the exact same application directly with GPT-5.6 Sol consumed just 9% of the limit. We saw the same trend with Fable 5. The subagent session cost $8.84, whereas the direct single-model build cost only $8.00. Crucially, the direct Fable 5 build took a mere 13 minutes—nearly cutting the development time in half compared to the collaborative approach. Write the Code Yourself, Senior This experiment exposes a clear parallel to human team dynamics. For a senior developer, writing the code directly is almost always faster than managing junior engineers, explaining requirements, and correcting their mistakes. While subagents make sense for highly repetitive, decoupled tasks, using them for integrated software development is currently a waste of time and resources. For now, skip the managerial overhead. Let your frontier models write the code directly.