58 lines
4.2 KiB
Markdown
58 lines
4.2 KiB
Markdown
# Multi-Agent System Reliability
|
|
|
|
## Metadata
|
|
- **Date:** 2026-04-13
|
|
- **Source:** https://blog.alexewerlof.com/p/multi-agent-system-reliability
|
|
- **Author:** Alex Ewerlöf
|
|
- **Category:** AI/Agent Architecture
|
|
|
|
## Key Insights
|
|
- LLMs are slow, error-prone, and stochastic — multi-agent topologies can propagate errors to the point of being useless
|
|
- Stop treating LLMs like "magic chatbots" — treat them as unreliable components in a distributed system
|
|
- Don't anthropomorphize LLMs — they have no fear of death, no empathy, and can't be motivated by threats
|
|
- 4 architecture patterns improve reliability: Hierarchy, Consensus, Adversarial Debate, and Knock-out
|
|
- Force correctness through architecture, not through emotional prompts or threats
|
|
- We need AI that is constrained, verified, pruned, and challenged — not AI that "cares"
|
|
|
|
## Summary
|
|
Multi-agent systems divide work across parallel and/or specialist agents to overcome LLM limitations like slowness and genericness. However, the underlying LLM remains unreliable (hallucination, logical fallacies, context drift), and multi-agent topologies can propagate these errors throughout the system.
|
|
|
|
This article presents 4 architecture patterns from human systems adapted for LLM reliability:
|
|
1. **Hierarchy** — A supervisor plans, breaks down tasks, distributes to workers, and validates results
|
|
2. **Consensus** — Multiple models vote; truth emerges from majority (3 models reduce same-hallucination probability from 20% to 0.8%)
|
|
3. **Adversarial Debate** — One agent proposes, another attacks, a judge moderates; prevents sycophancy
|
|
4. **Knock-out** — Multiple agents work on tasks, worst performers eliminated (cattle, not pets)
|
|
|
|
The core principle: don't ask models to "be careful" — force correctness through architectural constraints.
|
|
|
|
## Key Entities
|
|
- **Alex Ewerlöf** — Author, Senior Staff Engineer with 27 years experience, MS in Systems Engineering from KTH, SRE background, specializing in LLMs since 2023
|
|
- **Planner** — Smart model (e.g., Opus) that breaks user goals into small steps and distributes to workers
|
|
- **Worker** — Specialized agents (often smaller, faster models) that do one thing well
|
|
- **Validator** — Checkpoint that validates worker output; can be deterministic code or an LLM
|
|
- **Generator** — In adversarial debate, proposes initial ideas/plans
|
|
- **Critic** — Devil's advocate that attacks the generator's proposals
|
|
- **Judge** — Moderator that decides if critic is right and forces fixes
|
|
- **Watchdog** — Deterministic code pattern that breaks debate loops when thresholds are exceeded
|
|
|
|
## Key Concepts
|
|
- **Multi-Agent Hierarchy** — Supervisor pattern: Planner → Worker → Validator; dependency graph forces collaboration
|
|
- **Multi-Agent Consensus** — Majority voting across N models to cancel out individual noise and hallucinations
|
|
- **Multi-Agent Adversarial Debate** — Courtroom pattern preventing sycophancy; truth survives through opposition
|
|
- **Multi-Agent Knock-out** — Evolutionary selection; worst agents eliminated, survivors' traits combined
|
|
- **LLM Reliability Engineering** — Applying SRE principles to LLM systems; treating LLMs as unreliable components
|
|
- **Sycophancy** — Tendency of LLMs to please/agree even by lying when pressured with threats
|
|
- **Hallucination** — LLM generating false or invented information
|
|
- **Context Drift** — LLM losing focus or veering off topic during long interactions
|
|
- **Genetic Algorithms** — ML technique referenced by Knock-out pattern; fitness function evaluates solutions
|
|
- **Groupthink** — Can skew consensus results if agents have feedback loops between them
|
|
- **Bandwagon Effect** — Can skew consensus results; agents should run like a blind experiment
|
|
- **Cattle vs Pets** — SRE principle: treat LLM agents as replaceable "cattle," not unique beloved individuals
|
|
- **Dependency Graph** — Mechanism that forces model collaboration in Hierarchy pattern
|
|
|
|
## Related Sources
|
|
- [Multi-Agent Hierarchy](wiki/concepts/Multi-Agent-Hierarchy.md)
|
|
- [Multi-Agent Consensus](wiki/concepts/Multi-Agent-Consensus.md)
|
|
- [Multi-Agent Adversarial Debate](wiki/concepts/Multi-Agent-Adversarial-Debate.md)
|
|
- [Multi-Agent Knock-out](wiki/concepts/Multi-Agent-Knock-out.md)
|
|
- [Alex Ewerlof](wiki/entities/Alex-Ewerlof.md) |