Skip to content

Latest commit

 

History

History
51 lines (36 loc) · 2.28 KB

File metadata and controls

51 lines (36 loc) · 2.28 KB

Technical Challenges

1. API Cost vs. Latency

Problem: Running 200+ agents with premium LLMs generates massive API costs. Reducing model quality degrades simulation coherence.

Strategy:

  • Tiered model allocation — premium models only for critical decisions
  • Aggressive caching of persona definitions and frequent queries
  • Batch processing for non-time-sensitive operations (wiki updates, summaries)
  • Budget caps with automatic tier downgrade when limits approach

2. The Consensus Trap

Problem: LLM agents tend to converge on similar opinions, creating an echo chamber instead of diverse discourse.

Strategy:

  • Chaos Joker agents with explicit contrarian instructions
  • Entropy injection — random events that disrupt established consensus
  • Belief anchoring — each agent has immutable core beliefs that resist social pressure
  • Faction mechanics — structural opposition between groups

3. Managed Hallucination

Problem: LLM hallucinations are normally a bug. In NULL, they are a feature — but they must be managed.

Strategy:

  • Canon vs. Legends system — facts require multi-agent consensus to become Canon; unverified claims are tagged as Legends
  • Source tracking — every wiki claim links back to the conversation that generated it
  • Contradiction detection — automated flagging when new claims conflict with established Canon
  • Hallucination budget — configurable tolerance for creative vs. factual content per scenario

4. Context Window Management

Problem: Long-running simulations generate more context than any LLM can process in a single call.

Strategy:

  • Sliding window with periodic summarization
  • Hierarchical memory (short-term conversation, mid-term episode, long-term world state)
  • Vector-based retrieval for relevant historical context
  • Agent-specific memory pruning based on relevance

5. Emergent Behavior Stability

Problem: Multi-agent systems can enter degenerate states (infinite loops, deadlocks, runaway consensus).

Strategy:

  • Circuit breakers for repetitive conversation patterns
  • Diversity metrics with automatic intervention when thresholds are breached
  • Simulation health monitoring with automatic parameter adjustment
  • Periodic "epoch resets" that introduce structural changes