{
  "id": 991463,
  "title": "I Built a Multi-Agent Coding Orchestrator. It Kept Choosing Zero Workers.",
  "url": "https://urgent.news/2026/08/15/i-built-a-multi-agent-coding-orchestrator-it-kept-choosing-zero",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-15T07:47:40.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mahadansar/i-built-a-multi-agent-coding-orchestrator-it-kept-choosing-zero-workers-4bc3"
  },
  "original_language": "en",
  "account": "I built an AI coding orchestrator named Sol-Luna Orchestrator. The project started with the question of whether one powerful AI could decide when to delegate coding tasks to other AI agents. I created an open-source layer for OpenAI Codex that acts as a supervisor (GPT-5.6 Sol) and bounded workers (GPT-5.6 Luna instances). Sol Supervisor is responsible for overall task, verification, and final review, while Luna instances act as workers when Sol decides delegation is useful.\n\nThe orchestrator has two adaptive decisions: 1. Adaptive delegation, where Sol decides whether to delegate at all, and 2. Adaptive worker effort, where Sol decides how much reasoning effort each Luna worker needs. The goal was to let the strongest model decide how the work should be executed, not just to spawn more agents.\n\nParallel workers run in isolated Git worktrees, and tasks declare their intended file scope. Verification is independently rerun instead of trusting a worker's own PASS result. Workers cannot recursively invoke the orchestrator and create their own worker trees.\n\nTo determine when parallelism becomes faster, I designed progressively larger deterministic engineering fixtures with independent streams of work. The benchmark included four-module, six-module, and coupled control fixtures. In the six-module fixture, roughly 530 lines of specification and 85 deterministic assertions were used. Each module could be worked on independently, so six workers could theoretically make progress at the same time.\n\nThe benchmark compared three modes: Sol working alone (delegation disabled), free choice (delegation available but Sol decides), and forced parallel delegation (Sol required to delegate). Across all six free-choice runs, Sol declined to delegate, even when given the option to use the workers. When forced to delegate, Sol's execution was slower than the solo approach, with forced parallel execution being about 46% slower at four streams and 108% slower at six streams. Token usage also showed that forced parallel execution used significantly more tokens than solo execution.",
  "summary": "I expected more AI agents to make coding faster. They didn’t. That was not the result I was looking for, but it ended up being the most interesting result of the project. Over the past few weeks, I have been building Sol-Luna Orchestrator , an open-source orchestration layer for OpenAI Codex. The idea started with a simple question: What if one strong AI could decide when it actually needed help…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}