​
Strategy Routing

The Strategy Router classifies every incoming task and selects one of four execution strategies: SimpleReAct, DAG Parallel, MCTS, or Hypothesis. This ensures simple tasks run quickly, complex pipelines get full multi-agent execution, and open-ended scientific questions get iterative hypothesis-driven reasoning.

​
Why Strategy Routing?

Biological tasks vary widely:
  • Simple: “What does BRCA1 do?” or “Convert FASTA to GenBank.”
  • Structured: “Run QC, normalization, clustering, and annotation on pbmc.h5ad.”
  • Exploratory: “Find drug repurposing candidates for target X.”
  • Scientific hypothesis: “What causes immunotherapy resistance in melanoma?” or “Propose a mechanism for FOXP3-mediated T-cell suppression.”
A single pipeline for all would be slow for simple queries and insufficient for scientific reasoning. The router assigns each task to the most appropriate engine.

​
Two-Phase Classification

​
Phase 1: Heuristic Classification

Regex and structural cues score the task for each strategy:
  • Simple indicators: lookups, conversions, single-entity queries → boost SimpleReAct.
  • Pipeline indicators: multi-step keywords, commas/semicolons, step counts → boost DAG.
  • Exploratory indicators: discovery, candidate finding, screening → boost MCTS.
  • Hypothesis indicators: “what causes”, “propose mechanism”, “generate hypothesis”, “novel biomarker”, “how does X resist/evade” → boost Hypothesis.
Task length and structure (e.g. comma count) contribute bonus terms. Hypothesis strategy requires a minimum score of 2 and dominance over other categories. Confidence is computed as the max score over the sum of all scores (plus ε).

​
Phase 2: LLM Classification (Ambiguous Cases)

When confidence from Phase 1 is below 0.75, the task is sent to the fast LLM with a structured prompt that asks for:
  • Strategy choice (react / dag_parallel / mcts / hypothesis)
  • Confidence (0.0–1.0)
  • Short reasoning
  • Complexity score and estimated step count
The heuristic result is provided as a hint. The LLM response is parsed to produce the final TaskClassification.

​
Domain Detection

Twelve biological domain patterns (genomics, pharmacology, ecology, neuroscience, cell biology, immunology, cancer biology, molecular biology, microbiology, plant biology, marine biology, bioimaging) are matched against the task to produce a multi-label domain classification. This is used for:
  • Downstream tool and KG subgraph selection
  • Episodic memory domain filtering
  • Reporting and analytics

​
The Four Strategies

​
Strategy 1: SimpleReAct

When: Simple or direct tasks (lookups, conversions, single-step data questions). How: An enhanced ReAct loop:
  1. LLM produces a short numbered plan with checkboxes.
  2. Execute one step; observe output.
  3. Update plan with [✓]/[✗]; repeat until done or max steps reached.
Benefit: Up to ~3× speedup over the full multi-agent pipeline for simple tasks.

​
Strategy 2: DAG Parallel

When: Structured multi-step workflows (typical bioinformatics pipelines). How:
  1. Planner decomposes the task into a TaskDAG (nodes = subtasks, edges = dependencies).
  2. Execution proceeds by topological levels; within each level, all ready nodes run in parallel (asyncio.gather()).
  3. Each node: Execute → Validate (Critic) → Retry if needed (reflection-guided repair, deep retry, or LATM).
  4. Synthesizer compiles results into a single final report.
Benefit: Parallelism and isolation of failures; full self-correction and provenance tracking.

​
Strategy 3: MCTS (Exploratory)

When: Open-ended tasks with high epistemic uncertainty where the optimal analysis path is unknown a priori (e.g. drug discovery, candidate screening, target identification). How:
  1. Select: UCB1-guided traversal from the root to a leaf or expandable node.
  2. Expand: LLM generates several candidate next actions.
  3. Simulate: LLM estimates the value (reward) of the path.
  4. Backpropagate: Update visit counts and average rewards along the path.
  5. After a fixed budget of iterations, extract the best path (most-visited children).
  6. Convert the path to a DAG and execute via the standard pipeline.
Benefit: Exploration of multiple reasoning paths instead of greedy sequential steps. Ideal when you don’t know which tools or analyses will be most informative.

​
Strategy 4: Hypothesis (Scientific Reasoning)

When: Open-ended scientific questions requiring iterative hypothesis generation, testing, and refinement. Examples:
  • “What causes resistance to immunotherapy in triple-negative breast cancer?”
  • “Propose a mechanism by which FOXP3 suppresses effector T-cell activation.”
  • “What are novel biomarkers for early-stage pancreatic cancer?”
  • “How does KRAS G12C mutation drive resistance to targeted therapy?”
How: A Generate → Test → Evaluate → Refine loop across multiple rounds:
PhaseAction
GenerateLLM produces 3–5 distinct, falsifiable candidate hypotheses (statement + mechanism + testable prediction).
DesignFor each active hypothesis, design a targeted test: code execution, literature search, knowledge-graph query, or pure reasoning.
ExecuteRun the test (sandbox code / PubMed+Semantic Scholar / KG / LLM reasoning).
EvaluateScore evidence polarity (supports / contradicts / neutral) and confidence. Contradicting evidence is penalized 1.5×.
UpdateMark hypotheses as supported, rejected, or converged (confidence ≥ threshold).
RefineAfter round 2+, surviving hypotheses are refined based on accumulated evidence to avoid repeating failed directions.
SynthesizeProduce a final scientific report: leading hypothesis + evidence trail + alternative hypotheses + recommended next steps.
Evidence scoring (net_confidence): 
net_confidence = 0.5 + (Σ evidence_for - 1.5 × Σ evidence_against) / (2 × total)
Converges when net_confidence ≥ 0.80 (default, configurable) after at least 2 rounds. Inspired by:
  • Biomni A1 — <think>/<execute>/<observe> checklist loop with [✓]/[✗] markers
  • PantheonOS Evolution — Analyzer → Mutator → Evaluator with failed-hypothesis history (avoid repeating rejected directions)
  • CellType CLI — EvidenceReasoner with weighted claim scoring and contradiction detection
Benefit: Structured scientific reasoning with explicit evidence accumulation, contradiction tracking, and a full audit trail — rather than a single LLM answer pass.

​
Hypothesis Trigger Keywords

The following patterns automatically route to the Hypothesis strategy (heuristic phase):
PatternExample
hypothesis (any context)“Generate a hypothesis about X”
propose mechanism”Propose a mechanism for FOXP3 suppression”
what causes”What causes drug resistance in NSCLC?”
could … cause”Could BRCA1 loss cause genomic instability?”
mechanism of/behind/for”Mechanism of cisplatin resistance”
why does/do … [disease/resistance]”Why do tumors evade immune checkpoint therapy?”
generate hypothesis”Generate hypotheses about aging”
novel biomarker”Identify novel biomarkers for early PDAC”
potential mechanism”What are potential mechanisms for X?”
how does … resist/evade/escape”How does GBM evade temozolomide?”
propose … target/candidate/pathway”Propose candidate drug targets for ALS”

​
Strategy Comparison

DimensionSimpleReActDAG ParallelMCTSHypothesis
Best forDirect Q&A, conversionsMulti-step pipelinesUnknown optimal pathScientific mechanism questions
PlanningChecklist planFull TaskDAGTree search pathRound-by-round design
ParallelismSequentialWithin-level parallelSequentialPer-hypothesis parallel
Self-correctionRetry loopReflection + deep retryPath re-selectionContradiction tracking + refinement
OutputCode result / answerSynthesis reportSynthesis reportEvidence-annotated scientific report
Token costLowMedium–HighHighMedium (bounded by rounds)

​
Forcing a Strategy

Override auto-routing from the CLI or API: 
# CLI flag
biocortex run -s hypothesis "What drives EMT in pancreatic cancer?"
biocortex cli  # then: /strategy hypothesis

# Python API
agent.go(task, strategy=ReasoningStrategy.HYPOTHESIS)

​
Configuration

Strategy selection thresholds and LLM usage can be influenced by config (e.g. confidence threshold for LLM fallback, Hypothesis convergence threshold, max rounds). See Configuration for relevant options.

​
Next Steps