deep-research-swarm

Deep Research

Orchestrate multi-agent epistemic triangulation: diverge across research dimensions, detect overlaps and contradictions, verify deeply, then converge into a validated synthesis. Swarm parallelism serves epistemic robustness — not merely speed.

Adaptive routing ensures the pipeline fits the task: wide-search topics get a two-stage parallel swarm (breadth then depth); file-based tasks skip or augment external search; focused queries go straight to dimension decomposition.

Output Directory — MANDATORY

All deep research output files MUST be saved under:

/mnt/agents/output/research/

This is non-negotiable. Every file produced in any phase MUST use this directory as the base path. Do NOT save any research artifact to /mnt/agents/output/ directly — always use the /mnt/agents/output/research/ subdirectory.

Before writing any file, ensure the directory exists (create it if not).

Workflow Overview

User Query
  │
  ▼
Phase 0: Intent & Input Router
  │
  ├─ Route A: Wide Search (broad/exploratory, no clear dimensions)
  │   → Phase 1 (Quick Landscape)
  │     → Phase 1W (Multi-Agent Wide Exploration) ★ NEW
  │       → Phase 2 (Decompose, informed by rich landscape)
  │         → Phase 3 (Parallel Deep Dive)
  │           → Phase 4 (Cross-Verify) → Phase 5 (if conflicts)
  │             → Phase 6 (Insight Extraction) → Phase 7 (Report via writing skill)
  │
  ├─ Route B: Focused Search (specific question, clear dimensions)
  │   → Phase 1 (Landscape) → Phase 2 (Decompose)
  │     → Phase 3 (Parallel Deep Dive) → Phase 4 (Cross-Verify)
  │       → Phase 5 (if conflicts) → Phase 6 (Insight Extraction) → Phase 7 (Report)
  │
  ├─ Route C: File-Only Research (user explicitly restricts to file content)
  │   → Phase F (File Intake & Deep Analysis) ★ NEW
  │     → Phase 2 (Decompose from file themes)
  │       → Phase 3-F (Multi-Agent File Deep Dive, NO external search)
  │         → Phase 4 (Cross-Verify across file analyses)
  │           → Phase 6 (Insight Extraction) → Phase 7 (Report via writing skill)
  │
  └─ Route D: File-Augmented Research (files as primary reference + external supplement)
      → Phase F (File Intake & Deep Analysis) ★ NEW
        → Phase 1 (Targeted Landscape, informed by file gaps)
          → Phase 2 (Decompose, merging file themes + external landscape)
            → Phase 3 (Parallel Deep Dive, each agent has file context + search)
              → Phase 4 (Cross-Verify) → Phase 5 (if conflicts)
                → Phase 6 (Insight Extraction) → Phase 7 (Report via writing skill)