Research delegation heuristics guide the critical decision of when to deploy single-agent versus multi-agent exploration, how to decompose complex questions into parallelizable subtopics, and how to constrain exploration to prevent runaway resource consumption. These heuristics form the strategic layer of Research Workflow Architecture, determining how research execution scales with question complexity.
The delegation decision shapes everything downstream - token budgets, wall-clock latency, context management, and synthesis complexity. Poor delegation wastes resources through over-engineering simple queries or under-resourcing complex investigations. Effective heuristics match research strategy to question characteristics.
The Delegation Decision
Single Agent for Simple Queries: Straightforward fact-finding, lists and rankings, or queries with clear answers benefit from direct execution. “What are the top machine learning frameworks?” or “When was the transformer architecture introduced?” don’t require parallel exploration. A single agent can search, synthesize findings, and report efficiently.
Single-agent execution minimizes coordination overhead. No supervisor needed to decompose work. No synthesis required to combine findings. No Isolating Context needed to prevent contamination. The agent operates in one focused context from research brief to final findings.
Multiple Agents for Complex Investigations: Comparative analyses comparing multiple dimensions (X vs Y across criteria A, B, C), multi-faceted questions exploring different aspects of a topic, and PhD-level research requiring depth across subtopics all benefit from parallel agent deployment.
“Compare React, Vue, and Angular for enterprise applications” naturally decomposes into framework-specific exploration - one agent deep-dives on React’s enterprise characteristics, another on Vue, a third on Angular. Each agent develops expertise in its assigned domain without distraction from the others. The supervisor synthesizes these specialized insights into comparative understanding.
The multi-agent approach mirrors Amdahl’s Law - the theoretical speedup from parallelization depends on what proportion of work can be parallelized. Research often has high parallelizable proportion - distinct subtopics can be investigated independently. This makes multi-agent delegation particularly effective for complex queries.
Topic Decomposition Patterns
Effective decomposition creates non-overlapping subtopics with clear boundaries and complete coverage:
Dimensional Decomposition: Break multi-dimensional questions along their natural axes. “Research AI safety” might decompose into technical safety (alignment, robustness), governance safety (policy, regulation), and deployment safety (real-world risks, mitigation strategies). Each dimension explores a distinct facet without overlap.
Comparative Decomposition: For comparison queries, assign each comparison target to its own agent. “Compare SQL vs NoSQL databases” becomes two subtopics: deep investigation of SQL characteristics, deep investigation of NoSQL characteristics. The supervisor synthesizes comparative insights rather than each agent attempting comparison.
Chronological Decomposition: Historical questions might decompose by time period. “Evolution of programming paradigms” could split into procedural era (1950s-1970s), object-oriented era (1980s-2000s), and functional/concurrent era (2000s-present). Each agent develops deep understanding of its period.
Scale Decomposition: Questions spanning different scales split naturally. “Impact of social media” might decompose into individual-level effects (psychology, behavior), community-level effects (group dynamics, information spread), and societal-level effects (politics, culture). Each scale represents coherent investigation.
The key principle: boundaries should be clear enough that workers can operate independently without needing to coordinate or reference each other’s work. Overlapping subtopics create redundancy. Gaps in coverage miss important aspects. Clean decomposition prevents both failure modes.
Concurrent Agent Limits
Unbounded parallelization seems attractive but creates practical problems:
Context Multiplication: Each additional agent requires its own context window. Five agents with 50k token contexts consume 250k tokens concurrently. Ten agents would consume 500k tokens. Context costs scale linearly with agent count while quality gains may not.
Synthesis Complexity: The supervisor must synthesize findings from all workers. Combining insights from three agents is manageable. Combining insights from ten agents creates significant integration challenges. The supervisor’s context fills with worker outputs, potentially causing Context Distraction.
Coordination Overhead: More agents mean more management burden. The supervisor must ensure all agents stay on topic, handle failures or timeouts, and orchestrate completion. This overhead can dominate actual research time when agent counts grow large.
Anthropic’s implementation uses 3-5 concurrent agents as a practical maximum. This provides meaningful parallelization (3-5x speedup over single agent) while keeping synthesis tractable and context costs manageable. The exact limit depends on question complexity and available context budget.
Delegation as Resource Management
The supervisor should “think like a research manager with limited time and resources.” This framing emphasizes constraint-aware decision making:
Budget-Aware Decomposition: Given fixed token budget, how should resources distribute across subtopics? High-priority areas deserve more thorough investigation. Peripheral topics might receive surface-level treatment or exclusion entirely.
Time-Bounded Exploration: Research execution has time limits. Deploying many agents consumes wall-clock time through coordination overhead. Fewer agents with deeper exploration might finish faster than many agents with shallow passes requiring multiple synthesis rounds.
Quality Thresholds: Not every subtopic deserves multi-agent treatment. The delegation heuristic should identify which subtopics are critical enough to warrant dedicated agents versus which can be handled through single-agent broad strokes.
This connects to Reversible Decisions - the choice to deploy multiple agents should be high-confidence since coordination overhead makes reversal expensive. If uncertain whether multi-agent approach is necessary, consider starting with single agent then escalating to multi-agent if initial findings reveal complexity.
Iteration Limits
Research can potentially continue indefinitely - every answer reveals new questions. Hard iteration limits prevent runaway exploration:
Fixed Iteration Count: Simple approach limits research to N rounds. First round deploys agents based on research brief. If findings reveal gaps, second round addresses those gaps. After N rounds, research concludes regardless of remaining uncertainties.
Token Budget Limits: Alternative approach sets token budget cap. Research continues until token consumption exceeds threshold. This naturally adapts to question complexity - simple queries resolve quickly with few tokens, complex queries consume more budget.
Quality Thresholds: More sophisticated approach evaluates whether findings satisfy research objectives. If synthesis meets quality criteria, research concludes. Otherwise, additional rounds investigate deficiencies until quality threshold met or iteration limit reached.
The iteration limit acts as safety mechanism preventing infinite loops where research generates new research questions indefinitely. It enforces decisive completion even when perfect coverage remains unachieved.
Prompt Patterns for Delegation
The research supervisor needs prompts that guide systematic delegation decisions:
You are a research supervisor managing a team of specialized research agents.
Given the research brief, decide:
1. Single agent (simple query) or multiple agents (complex investigation)?
2. If multiple agents:
- How many agents? (Max 5 for tractability)
- How to decompose into non-overlapping subtopics?
- What specific question should each agent investigate?
Delegation criteria:
- Single agent for: fact-finding, lists, straightforward queries
- Multiple agents for: comparisons, multi-faceted analysis, PhD-level depth
Decomposition principles:
- Clear boundaries (no overlap between agents)
- Complete coverage (no gaps in research)
- Balanced scope (similar complexity across agents)
Output a structured delegation plan specifying agent count and
assigned subtopics.
This prompt establishes supervisor role, defines decision points (single vs multi, decomposition strategy), provides decision criteria, and constrains output format. The pattern follows Prompt Engineering principles of clarity, structure, and explicit guidance.
Integration with Orchestrator-Worker Pattern
Delegation heuristics implement the Orchestrator-Worker Pattern strategically. The supervisor orchestrates, but how it orchestrates depends on delegation decisions:
No Delegation: Supervisor becomes the worker, directly executing research without spawning sub-agents. This collapses the orchestrator-worker distinction for simple cases.
Fixed Delegation: Supervisor spawns predetermined number of workers based on question structure. Comparative queries always get N agents (one per comparison target). This provides predictable resource usage.
Dynamic Delegation: Supervisor analyzes brief and determines optimal agent count at runtime. Complex questions spawn more agents, simple questions fewer. This adapts resource allocation to actual complexity.
The heuristics guide this orchestration, translating from abstract research brief to concrete agent deployment strategy. Poor heuristics create inefficient resource allocation - too many agents for simple queries, too few for complex investigations.
The OODA Connection
Delegation reflects the OODA Loop’s orient-decide-act cycle. The supervisor observes the research brief, orients by analyzing complexity and decomposition opportunities, decides on single versus multi-agent approach, and acts by deploying agents accordingly.
Effective delegation requires rapid orientation - understanding question structure quickly enough to make deployment decisions without extensive analysis paralysis. The heuristics encode pattern recognition that enables this rapid orientation. “Comparative query → multi-agent delegation” or “fact-finding → single agent” become automatic responses rather than deliberative decisions.
Learning from Execution
Initial heuristics will make mistakes - deploying too many agents for manageable questions or too few for complex investigations. Effective systems should learn from these failures:
Token-Quality Correlation: Track relationship between agent count and output quality. Do five-agent deployments consistently outperform three-agent for certain question types? Use this correlation to refine delegation heuristics.
Synthesis Difficulty Signals: If supervisor struggles to synthesize findings (requests clarification, identifies conflicts), it might indicate poor decomposition. The subtopics created overlap or gaps. This signals need for improved decomposition patterns.
Iteration Analysis: Questions requiring many iteration rounds might benefit from different initial delegation. If research consistently needs three rounds to achieve quality, perhaps initial agent count should increase.
These feedback loops transform static heuristics into adaptive strategies that improve through experience. The delegation decision becomes less about fixed rules and more about learned patterns across many research executions.
The Fundamental Tradeoff
Delegation balances comprehensiveness against efficiency. Single-agent research is token-efficient and simple to manage but may miss nuances. Multi-agent research provides thorough coverage but consumes resources through parallelization and coordination overhead.
The heuristics exist to navigate this tradeoff intelligently - deploying resources where complexity justifies costs, economizing where simplicity suffices. Getting delegation right means research that scales appropriately with question difficulty rather than applying fixed approaches regardless of need.