Fermi

synopsis

Fermi is a Probabilistic Reasoning Engine (PRE) designed to bring rigorous uncertainty quantification to agentic systems. Named after Enrico Fermi's approach to estimation under uncertainty, the engine provides a domain-specific language — the Forecasting Programming Language (FPL) — for expressing probabilistic models, defining confidence intervals, and composing Fermi-style estimations that decompose complex questions into tractable sub-problems with quantified uncertainty at every stage.

At its core, Fermi is the reasoning substrate that powers the Agent Bestiary's forecasting capabilities. Each agent in the bestiary uses FPL to structure its reasoning, calibrate its confidence, and produce evidence-weighted outputs. The engine handles distribution arithmetic, sensitivity analysis, gas-fee accounting for computational cost, and multi-model execution — enabling agents to reason probabilistically about everything from market dynamics to biological system behavior. Fermi transforms agents from black-box predictors into transparent, auditable reasoning machines.

key capabilities

• Forecasting Programming Language (FPL) — a purpose-built DSL for probabilistic reasoning with lexer, parser, AST, and evaluator
• Distribution arithmetic — native support for normal, log-normal, beta, uniform, and triangular distributions with algebraic composition
• Sensitivity analysis — automated identification of which input assumptions most affect output uncertainty
• Gas-fee accounting — computational cost tracking per reasoning step, enabling economic optimization of agent workloads
• Multi-model execution — orchestrate reasoning across multiple LLM backends with confidence-weighted aggregation
• Agent memory integration — episodic memory storage with embedding-based retrieval for context-aware forecasting
• Ontology-driven knowledge graphs — structured domain knowledge accessible to agents during reasoning
• MCP server — Model Context Protocol integration for IDE-native access to Fermi agents via Zed and other editors

architecture

• FPL Engine — Rust-native lexer, parser, evaluator, and semantic analyzer for the Forecasting Programming Language
• Agent Backend — Axum-based API server with PostgreSQL persistence, agent registry, and execution framework
• Memory System — episodic memory store with Anthropic embeddings for semantic retrieval and consolidation
• Coherence Engine — self-optimizing loop that monitors agent output quality and adjusts reasoning parameters
• Projector — forward-looking projection engine with caching layer for trend extrapolation
• MCP Server — stdio-based Model Context Protocol server exposing agent tools to IDE environments

related research

• FPL: Agent-Assisted Architecture
• Agentic Infrastructure White Paper
• Coherence Improvement Loop
• Cross-Domain Swarm Intelligence

tags