---
title: Benchmarks
description: "70+ harnesses for agents, models, and embodied systems."
---

The benchmarks suite orchestrates evaluation runs across every Eliza subsystem. It covers tool-call correctness, agent autonomy, voice and vision, embodied control, long-horizon reasoning, and adversarial robustness.

## Categories

- **Eliza-1** — model-side: tool-calling, instruction following, agentic loops.
- **SWE-Bench** — software engineering tasks.
- **LifeOps-Bench** — long-running personal-assistant scenarios.
- **Orchestrator** — multi-agent coordination harness.
- **Vision-Language** — multimodal grounding.
- **Voice-Emotion** — speech recognition and affect under noise.
- **Feed / ScamBench** — adversarial social pressure inside the Feed simulation.
- **Robot** — embodied skill evaluation.

## Running a benchmark

The orchestrator at `packages/benchmarks/orchestrator/` accepts a config and runs the named harness, collecting traces and metrics. Configs pin the model under test, dataset shard, evaluation criteria, and reporter.

## Reports

Each run emits a calibration report, scorecards, and per-task traces. Reports are checked into `packages/benchmarks/reports/` for reference runs.

## Source

- Suite root: [`packages/benchmarks/`](https://github.com/elizaOS/eliza/tree/develop/packages/benchmarks)
- Orchestrator: `packages/benchmarks/orchestrator/`

## See also

- [Eliza-1](/tracks/training/eliza-1)
- [Feed](/tracks/training/feed)
- [Robot](/tracks/training/robot)