Prospective Refinement through Intelligent Synthesis and Multiplicity
Exploit native LLM stochasticity instead of engineering artificial personas.
The classic multi-agent mistake is creating artificial "roles" (The Expert, The Critic, The Devil's Advocate) that bias the model's latent space and produce caricatural, predictable responses.
Role-playing forces the model into a constrained subspace. You get verbose "experts" and contrarian "critics" by design — not by intelligence.
PRISM operates on a fundamentally different principle: Emergent Perspective Orchestration.
Inject the same prompt into N identical instances simultaneously.
Prompt P → [Agent₁, Agent₂, Agent₃, ... Agentₙ]
No roles. No constraints. Each agent explores a different probability trajectory through the model's latent space, driven purely by the temperature-based stochasticity native to LLMs.
You can use:
- Same model × N instances (maximum variance from same distribution)
- Different models (cross-architecture diversity)
- Both simultaneously
Each agent generates a unique solution based on the natural fluctuations of sampling. This reveals nuances, edge cases, and approaches that a single agent — or a role-constrained agent — would systematically ignore.
This is not a bug. It's the feature.
The Meta-Agent is not a vote counter. It acts as a Structural Architect:
- Identifies the most robust logical anchor across all responses
- Extracts critical variables identified by minority instances
- Fuses them into a single optimal output
This kills "Consensus Soup" — the mediocre averaged answer you get from naive majority voting.
[Response₁, Response₂, ... Responseₙ] → Meta-Agent → Optimal Output
| Persona Approach | PRISM | |
|---|---|---|
| Diversity source | Engineered role constraints | Native LLM stochasticity |
| Bias | High (role distortion) | Low (no artificial constraint) |
| Consistency | Predictable by design | Emergent |
| Synthesis | Majority vote / consensus | Meritocratic arbitrage |
| Architecture | Stateful, role-dependent | Stateless, domain-agnostic |
| Scalability | Hard (role management) | Easy (N is just a parameter) |
Anti-Bias — No role-playing distortion. The model reasons freely in its full latent space.
Self-Consistency — The goal isn't the majority answer. It's the technical optimum by data fusion.
Scalability — Stateless architecture. Domain-agnostic. N is just a number. Run it on fitness programming or legal document analysis — same protocol.
PRISM was originally developed for personalized fitness programming. Applied to biomechanics and physiology:
- N instances each generate a complete training plan from the same athlete profile
- Each plan reflects different periodization strategies, exercise selections, intensity distributions
- Meta-Agent identifies the most evidence-consistent structure and injects the high-value variations
Result: plans that cover edge cases a single model call would statistically miss.
import { PRISMSampler } from './examples/javascript/basic-sampling.js';
const prism = new PRISMSampler({
n: 5, // Number of parallel instances
temperature: 0.9,
model: 'claude-opus-4-6'
});
const result = await prism.run(
prompt,
metaAgentPrompt // See /prompts/templates/
);
console.log(result.synthesis); // The meritocratic output
console.log(result.divergences); // What each agent found uniquelyfrom examples.python.meta_agent import PRISMPipeline
pipeline = PRISMPipeline(n=5, model="claude-opus-4-6")
result = pipeline.run(prompt, meta_prompt)
print(result["synthesis"])
print(result["divergences"])prism-framework/
├── README.md
├── LICENSE
├── .env.example
├── package.json # Node.js dependencies
├── requirements.txt # Python dependencies
├── docs/
│ ├── architecture.md # Deep dive into the protocol
│ ├── vs-personas.md # Extended comparison with persona approaches
│ └── daco.md # DACO companion protocol spec
├── examples/
│ ├── javascript/
│ │ ├── basic-sampling.js # N-Sampling with Anthropic SDK
│ │ └── meta-agent.js # Full PRISM pipeline
│ └── python/
│ ├── basic_sampling.py
│ └── meta_agent.py
└── prompts/
└── templates/
├── master-synthesis.md # Meta-Agent prompt template
└── sampling-base.md # Base sampling prompt guidelines
PRISM's companion protocol for tool-augmented agents.
The pattern: the LLM poses the questions, not the developer. The agent receives a task description + a manifest of available MCPs. It builds the execution graph, resolves dependencies, fires parallel tool calls, and recovers from failures — autonomously.
First production implementation: Smart Rabbit MCP (npx smartrabbit-mcp) — Claude Desktop asks the user for their fitness profile in natural language, orchestrates the Smart Rabbit API + PubMed, and returns a complete program.
See docs/daco.md for the full specification.
Jacques Chauvin — WNBF World Champion (4th), fitness AI systems builder.
- Site: jc-coaching.com
- Framework in production: Smart Rabbit Fitness
- PRISM community: jc-coaching.com/prism
- immune — Adaptive memory system — learns patterns from every scan (+85% code quality)
- chimera — Bio-inspired 3-stage pipeline (Slime Mold → PRISM → Immune)
- spinal-loop — Neuromuscular-inspired agent routing (cheap models first)
- daco-framework — Declarative Agent & MCP Orchestration on Cloudflare Workers
- smartrabbit-mcp — AI workout generator MCP server (smartrabbitfitness.com)
MIT — use it, fork it, build on it. Attribution appreciated, not required.