Multi-Agent Research Orchestration Patterns

Full reference on proven patterns for multi-agent research. Extracted from Anthropic's engineering blog, production systems (GPT-Researcher, Onyx, LangChain), and 40+ sources.

For the comprehensive research document with all sources, see: /mnt/c/Users/olled/Documents/Obsidian/Notes/02 - Content/Research/Multi-Agent Research Orchestration Patterns 2026.md

Key Statistics

Proven Architecture: Orchestrator-Worker

Every production system converges on this:

``` User Query | v [Lead Agent / Orchestrator] | ├── Clarify (if needed) ├── MECE decompose into sub-questions ├── Spawn 3-6 parallel sub-agents │ ├── Sub-agent A (angle 1) → findings │ ├── Sub-agent B (angle 2) → findings │ └── Sub-agent C (angle 3) → findings ├── Evaluate coverage → spawn more if gaps ├── Synthesize all findings └── Generate final report with citations ```

Anthropic uses: Opus 4 lead + Sonnet 4 workers. Different models reduce single-model blind spots.

The 10 Pitfalls to Engineer Against

1. Agent spawning explosions — Hard-cap at 6 agents. Scale to complexity.
2. Duplicate work — MECE + explicit "do NOT cover X" boundaries.
3. Overly-specific search queries — Explicitly prompt "start broad (2-4 words), then narrow."
4. SEO content farm preference — Add source quality heuristics to every prompt.
5. Context exhaustion — Sub-agents write to files, lead reads summaries.
6. Anchoring bias — Batch ALL results before synthesis. Never process sequentially.
7. Lost nuance in compression — Preserve: facts, numbers, quotes, citations, contradictions.
8. "Good enough" stopping — Require 2+ sources per sub-question before report.
9. Excessive inter-agent chatter — Sub-agents report only to lead. No broadcasting.
10. Single model bias — Use Opus for lead, Sonnet for workers when possible.

Decomposition Strategies

MECE (Mutually Exclusive, Collectively Exhaustive)

The gold standard. Each sub-question covers a distinct area. No overlaps, no gaps.

Perspective-Based

Break by stakeholder/lens: technical, business, user experience, historical, competitive, future.

Dimension-Based

Break by question type: what exists, what are problems, what solutions emerge, who are key players, what does evidence say.

Scale-Effort-to-Complexity

• Simple fact-finding: 1 agent, 3-10 tool calls
• Direct comparisons: 2-4 agents, 10-15 calls each
• Complex multi-faceted: 5-6 agents with divided responsibilities

Synthesis Best Practices

Three-Stage Pipeline

1. Aggregate: Collect all findings. Identify thematic connections.
2. Cross-reference: Where do sources agree (high confidence)? Where do they contradict?
3. Construct narrative: Organize by theme, not by agent. The output should feel like one coherent document, not a concatenation.

The Critique Loop (Optional)

After synthesis, optionally spawn a critic agent that checks:

• Completeness (all sub-agent findings reflected?)
• Accuracy (claims properly attributed?)
• Coherence (logical flow?)
• Gaps (what remains unanswered?)

If issues found, revise synthesis.

Production Systems to Study

1. Anthropic's multi-agent system — Opus lead + Sonnet workers, 90.2% improvement
2. GPT-Researcher (github.com/assafelovic/gpt-researcher) — 14k+ stars, editor/researcher/writer pipeline
3. Onyx (github.com/onyx-dot-app/onyx) — Enterprise-grade, two-tier orchestration
4. LangChain Open Deep Research (github.com/langchain-ai/open_deep_research) — StateGraph workflow
5. Google ADK — ParallelAgent → SequentialAgent pattern