research-orchestrator

A Claude Code skill that runs deep research as a multi-agent pipeline instead of a single long conversation. Splits a research question into parallel investigations, lets the agents share notes through shared memory, then synthesizes and judges the result.

Architecture

Splitter → Round 1 (N parallel agents) → Coordinator (shared memory)
        → Round 2 (gap-filling agents) → Synthesizer → Judge

• Splitter breaks the question into independent sub-questions.
• Round 1 agents investigate in parallel, writing findings to a shared memory store.
• Coordinator reads the round-1 outputs, identifies gaps and contradictions.
• Round 2 agents target the gaps the coordinator flagged.
• Synthesizer produces the final report.
• Judge scores the output against the original brief.

Status

• v3 built and deployed as a Claude Code skill.
• Repo: LucianoLupo/research-orchestrator
• Skill: ~/.claude/skills/research-orchestrator/SKILL.md

Test results

• Quality: 8.0–8.2/10 on multi-source research benchmarks (judged by an independent agent against the brief).
• Cost: ~$13 for a standard run, ~$5 for a quick run.

Why I built it

A single Claude Code conversation researching a hard topic loses depth as the context fills with web fetches and intermediate reasoning. Splitting research across parallel agents that each get a fresh context window — with one coordinator mediating — produces measurably better synthesis at lower per-token cost than scaling a single context.

Connection points

• claude-autoresearch — the milestone-driven autonomous version of this same pattern.
• agent-orchestrator — the runtime that supervises long-running multi-agent workflows.
• Production LLM Eval Platforms research — produced by an earlier version of this pipeline.