Dynamic Agent Selection: How PAK4L Chooses Its Experts

A privacy policy and a construction tender have almost nothing in common. They live in different regulatory universes, use different vocabulary, and require completely different expertise to evaluate. Sending the same set of agents to review both would be like sending an orthopedic surgeon to perform cardiac surgery — technically they're both doctors, but you want a specialist.

PAK4L's DynamicAgentFactory solves this by assembling a custom review panel for every document.

How It Works

When a document arrives, the system goes through a three-stage process:

1. Document Classification

First, a fast LLM (Gemini Flash) reads the first ~2000 characters and classifies the document into one of several categories: contract, privacy policy, procurement proposal, corporate bylaw, legislation, technical specification, and more. This classification is fast — typically under 2 seconds.

2. Agent Selection

Based on the classification, the factory selects agents from a pool of 40+ available specialists. The selection follows a guidance map — a mapping from document categories to recommended agent sets:

# Simplified guidance map
guidance_map = {
  "contract": [
    "logic_checker", "completeness_auditor",
    "italian_eu_law_expert", "consistency_checker"
  ],
  "privacy_policy": [
    "gdpr_compliance", "ethics_reviewer",
    "constitutional_analyst", "regulatory_compliance"
  ],
  "procurement": [
    "procurement_technical_reviewer",
    "architecture_reviewer", "ip_expert",
    "financial_analyst", "regulatory_compliance"
  ]
}

Every document also gets a set of core agents that run regardless of type: `clarity_expert`, `style_editor`, `structure_analyst`, `fact_checker`, and `audience_alignment`. These ensure baseline quality coverage.

3. Tier-Based Limits

The number of active agents is gated by the user's subscription tier. Free users get the core agents plus a limited number of specialists. Professional and Enterprise users get the full panel. This ensures fair resource allocation without compromising the quality of core analysis.

Batched Execution

Agents don't all run simultaneously. The orchestrator launches them in batched waves — typically 4-6 agents per wave. This balances speed (parallelism) with resource management (API rate limits). Each wave completes in 15-30 seconds, and a typical review finishes in 3-5 minutes.

Custom Agents

Beyond the built-in pool, users can create custom agents with specific personas and evaluation criteria. A law firm might create a "M&A Due Diligence" agent. A compliance department might build a "SOX 404 Auditor". These custom agents participate in the same pipeline — dynamically selected when relevant, running alongside the standard specialists.