multiagent-eval

Propagation-aware evaluation for multi-agent AI systems.

Single-LLM eval tools (RAGAS, DeepEval) miss what actually breaks in production: errors that start in Agent 1, silently propagate through Agent 3, and surface in the final output with no trace of origin.

multiagent-eval finds where the fault began.

---

The Problem Nobody Is Solving

Agent 1 ──► Agent 2 ──► Agent 3 ──► Agent 4 (Writer)
   │           │           │            │
   ✓           ✗           ✗            ✗  ← Error propagates; eval sees only final failure

You run eval. Score looks good. You ship.

Three days later: a hallucination in production. You check your eval results. Everything passed.

What happened?

Your eval checked the final output. It didn't check whether Agent 2 silently corrupted the information Agent 1 found. It didn't check whether Agent 3's hallucination was its own fault, or the result of broken input from upstream.

That's the gap multiagent-eval closes.

---

Quickstart

git clone https://github.com/iremsusavas/multiagent-eval.git
cd multiagent-eval
pip install -e .

# Zero-dependency demo — no API key needed
python examples/quickstart_mock.py

LLM-based evaluation requires a running LLM. Supports OpenAI, Anthropic, or local models via Ollama (no API key needed):

ollama pull llama3.2

Then in eval_config.yaml:

judge:
  primary_model: "ollama/llama3.2"
  api_base: "http://localhost:11434"

For a fully zero-dependency demo (no LLM needed):

python examples/quickstart_mock.py

---

What Makes This Different

Propagation Judge

Detects where information corruption begins — not just that it happened. Builds a directed graph where each edge carries a fidelity score. Red edges show exactly where data was lost or distorted between agents.

Built-in Bias Detection

Every LLM judge call automatically runs:

• Primacy bias (A/B swap permutation tests)
• Verbosity bias (length vs. correctness)
• Tone bias (neutral vs. apologetic framing)
• Cascade bias (upstream error penalizing innocent agents)

CI/CD Native

Eval isn't a report. It's a gate.

# .github/workflows/multiagent-eval.yml
- name: Run evaluation
  run: multiagent-eval run --config eval_config.yaml
# eval_score < threshold → fail the PR

Statistical Rigor

Bootstrap confidence intervals and permutation p-values on every run. "Did we improve?" becomes answerable.

Failure Mode Taxonomy

Not just a score. A category:

PROPAGATION_ERROR | HALLUCINATION | CONTEXT_LOSS | ORCHESTRATION_BREAK | CASCADE_FAILURE | PII_LEAKAGE

---

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│                         multiagent-eval                                  │
├─────────────────────────────────────────────────────────────────────────┤
│  core/           trace, metrics, runner, state_machine, LLMGateway        │
│  judges/         LLMJudge (CoT, bias), ConsistencyJudge, PropagationJudge│
│  bias_detection/ primacy, verbosity, tone, cascade                       │
│  golden_datasets/ schema, manager, annotator, inter-rater agreement       │
│  reports/        JSON, HTML (D3.js), Streamlit dashboard                  │
│  integrations/   LangGraph, CrewAI, AutoGen, Custom adapters               │
│  telemetry/      OpenTelemetry spans → Datadog, Grafana, Jaeger          │
└─────────────────────────────────────────────────────────────────────────┘

---

Integrations

| LangGraph | CrewAI | AutoGen | Custom |

---

Production Features

• OpenTelemetry: Real-time span emission to Datadog/Grafana/Jaeger
• PII Detection: Email, SSN, credit card — zero-tolerance config
• Prompt Injection Detection: Pattern-based, extensible
• Cost Estimation: Know your budget before you run (estimate-cost --dataset ...)
• Regression Testing: Which examples degraded between v1.1 and v1.2? (regression-diff)

---

CLI

multiagent-eval run --config eval_config.yaml
multiagent-eval run --all                    # All examples in golden dataset
multiagent-eval estimate-cost -d datasets/research_qa.json
multiagent-eval regression-diff -a result_v1.json -b result_v2.json
multiagent-eval report --input results.json --format html
multiagent-eval dataset add --name my_dataset
multiagent-eval dashboard

---

Background

Built by an ML engineer who spent months improving LLM-as-Judge agreement from 63% to 84% in production at Pipedrive — and discovered that most eval problems aren't scoring problems. They're architectural ones.

JudgeGuard (primacy bias detection) came first. multiagent-eval is what came after asking: "What happens to these biases when you have five agents?"

---

Roadmap

• Leaderboard / MAE-Bench public benchmark
• Multi-turn stateful session evaluation
• Visual diff UI for agent output comparison
• Automated rubric improvement suggestions
• Native LangSmith integration

---

Contributing

Issues, PRs, and dataset contributions welcome. If you're building multi-agent systems and hitting eval problems — open an issue. That's how this gets better.

---

License

MIT