Autobench

When an application changes, you need to know whether the new version is more correct, faster, cheaper, or more reliable. Teams usually answer that question with a custom benchmark script: a small program that runs representative inputs, checks the outputs, collects measurements, and prints a comparison. As the application grows, these scripts are repeatedly rewritten and their results become difficult to reproduce or compare.

We built Autobench to solve that problem. Define the cases, variants, task, scores, and reports once in YAML or Python. Autobench runs the experiment, collects the evidence, records exactly what happened, and lets you replay, report, compare, or export the result without rebuilding that infrastructure for every application.

It is a YAML-first Python framework for AI and non-AI systems:

• deterministic dataset x variant execution
• sync and async application tasks
• semantic observations, checks, measurements, artifacts, and ABP traces
• built-in and custom scoring, cost derivation, policies, and paired baselines
• native Pydantic AI, OpenAI, OpenAI Agents, and HTTPX instrumentation
• explicit and automatic prompt/tool/schema/agent asset lineage
• immutable YAML records, replay, Rich reports, comparisons, and exports
• optional immutable-record export to OTLP-compatible telemetry backends

Install

uv add autobench

For native SDK instrumentation:

uv add 'autobench[instrumentation]'

For outbound OTLP HTTP/protobuf export:

uv add 'autobench[otlp]'

First Run

autobench validate examples/minimal/autobench.yaml
autobench run examples/minimal/autobench.yaml --record /tmp/autobench-minimal
autobench replay /tmp/autobench-minimal
autobench report /tmp/autobench-minimal

A task is a normal sync or async callable:

from autobench import Case, RunContext


def run(ctx: RunContext, case: Case) -> Result:
    mode = ctx.factor("mode")
    with ctx.span("subject", kind="workflow") as span:
        result = application(case.input, mode=mode)
        span.set_output(result)
        return result

The YAML spec owns reusable benchmark infrastructure: cases, variants, scoring, derivation, policies, instrumentation, and reports.

Examples

Run the offline release matrix:

make examples

Documentation

Full documentation: vcoderun.github.io/autobench

• Installation
• First Benchmark
• Use Cases
• Architecture
• YAML Spec
• Python API
• Autobench Protocol
• OTLP Export

LLM-readable indexes are available at llms.txt and llms-full.txt.

Development

uv sync --extra dev --extra instrumentation --extra openai-agents --extra otlp
make prod
make pre-commit

The release gate enforces Python 3.11-3.14, strict lint and typing, strict documentation builds, offline examples, and 100% source line and branch coverage.