credence-agents 0.2.0

Bayesian decision-theoretic agents — expected utility maximisation, not prompt engineering

Credence

Bayesian decision-theoretic agents — a reusable library and an empirical benchmark.

Credence provides domain-agnostic Bayesian agents that learn tool reliability from experience, use value-of-information calculations to decide when to query and when to abstain, and adapt when reliability changes. No heuristics, no hardcoded routing — just expected utility maximisation.

The project also includes a head-to-head benchmark comparing the Bayesian agent against LangChain ReAct agents on a multi-tool question-answering task.

Using Credence as a Library

# Core library (numpy + scipy only)
uv sync

# With benchmark dependencies (matplotlib, langchain, etc.)
uv sync --extra benchmark

# With dev tools (pytest, ruff, etc.)
uv sync --extra dev

Facade import

from credence import BayesianAgent, ToolConfig, ScoringRule

Minimal custom-domain example

import numpy as np
from credence import BayesianAgent, ToolConfig, ScoringRule

# Define your domain's categories and tools
categories = ("plot", "character", "setting", "theme")
num_cats = len(categories)
tools = [
    ToolConfig(cost=1.0, coverage_by_category=np.ones(num_cats)),   # wiki_search
    ToolConfig(cost=3.0, coverage_by_category=np.ones(num_cats)),   # deep_read
    ToolConfig(cost=2.0, coverage_by_category=np.ones(num_cats)),   # summary_llm
]

# Category inference: question text -> probability vector over categories
# (or use make_keyword_category_infer_fn for keyword-based heuristics)
def infer_category(text: str) -> np.ndarray:
    return np.ones(num_cats) / num_cats  # uniform prior

agent = BayesianAgent(
    tool_configs=tools,
    categories=categories,
    category_infer_fn=infer_category,
)

Categories, tools, scoring rules, and the category inference function are all injected — the agent itself is domain-agnostic.

The Benchmark

A head-to-head comparison between a Bayesian decision-theoretic agent and LangChain ReAct agents on a multi-tool question-answering task.

Both agents have access to the same four tools with heterogeneous, category-dependent reliability. The Bayesian agent learns tool reliability from experience and uses value-of-information calculations to decide when to query, when to cross-verify, and when to abstain. The LangChain agent lets the LLM decide.

Running experiments

# Run tests
uv run pytest -v

# Run all experiments (stationary + drift + ablation)
uv run python -m experiments.run_full_comparison

# Or run individually
uv run python -m experiments.run_stationary --seeds 20
uv run python -m experiments.run_drift --seeds 20
uv run python -m experiments.run_ablation --seeds 20

LangChain agents require a local Ollama instance with llama3.1 (default). Set CREDENCE_LLM_PROVIDER=openai or CREDENCE_LLM_PROVIDER=anthropic for API-based models.

Key Results

Agent	Score	Accuracy	Tools/Q
oracle	+188.2	70.6%	1.08
bayesian	+112.6	59.6%	0.99
langchain_react	-7.4	64.0%	3.22
langchain_enhanced	-68.2	66.0%	3.94

The Bayesian agent outscores LangChain by 120 points despite lower accuracy — it queries ~1 tool per question instead of ~3.2, and strategically abstains on low-confidence questions. Enhanced prompting makes LangChain worse by triggering more tool calls without proportional accuracy gains.

See results/RESULTS.md for full results across all three experiments (stationary, drift, ablation).

Project Structure

credence/
├── credence/                    # Facade package — public API re-exports
│   └── __init__.py
├── src/
│   ├── inference/               # Domain-agnostic Bayesian inference layer
│   │   ├── beta_posterior.py    # Beta-Bernoulli reliability tracking
│   │   ├── voi.py               # Value of information, ScoringRule, ToolConfig
│   │   └── decision.py          # EU-based decision logic
│   ├── agents/
│   │   ├── bayesian_agent.py    # Domain-agnostic Bayesian agent
│   │   ├── common.py            # Shared agent interface
│   │   ├── baselines.py         # Random, all-tools, oracle agents
│   │   ├── langchain_agent.py   # Standard LangChain ReAct agent
│   │   └── langchain_enhanced.py # LangChain with best-effort prompting
│   ├── environment/             # Benchmark-specific: simulated tools and questions
│   │   ├── benchmark.py         # The benchmark harness
│   │   ├── tools.py             # Simulated tools with known reliability
│   │   ├── questions.py         # Question bank with ground truth
│   │   └── categories.py        # Category definitions, make_keyword_category_infer_fn
│   ├── analysis/
│   │   ├── metrics.py           # Score, calibration, cost metrics
│   │   └── visualisation.py     # Plots and dashboards
│   └── utils/
│       └── logging.py           # Structured logging for analysis
├── tests/
├── experiments/
│   ├── run_stationary.py        # Main experiment: stationary reliabilities
│   ├── run_drift.py             # Extension: tool reliability drift
│   └── run_ablation.py          # Ablation studies
├── results/                     # Generated plots and data
└── pyproject.toml               # Core deps vs [benchmark] vs [dev] extras

Design Principles

1. Everything is expected utility maximisation — no heuristics
2. No hacks — if it doesn't work, fix the model
3. LLM outputs are data — with quantified uncertainty
4. The benchmark must be fair — LangChain gets every advantage
5. Be honest about parameters — every number is justified

See CLAUDE.md for architecture and development guidelines. See SPEC.md for the full mathematical specification.