Benchflow

BenchFlow — Agent Benchmarking

BenchFlow runs AI coding agents against tasks in sandboxed environments and scores their output via ACP (Agent Communication Protocol).

Arguments passed: $ARGUMENTS

Dispatch on arguments

No args or status — show current state

1. Check if benchflow is installed: uv tool list | grep benchflow
2. Check if API keys are set (GEMINI_API_KEY, ANTHROPIC_API_KEY, etc.)
3. Check available agents: bench agent list
4. Show recent eval results if any exist under jobs/ (the default --jobs-dir)
5. Point to next action based on state

run <task-path> — run a single task

bench eval run \
  --tasks-dir <task-path> \
  --agent gemini \
  --model gemini-3.1-flash-lite-preview \
  --sandbox daytona

Or via Python SDK:

import asyncio
import benchflow as bf
from benchflow import RolloutConfig, Scene
from benchflow._utils.benchmark_repos import resolve_source

async def main():
    config = RolloutConfig(
        task_path=resolve_source("benchflow-ai/skillsbench", path="tasks/edit-pdf"),
        scenes=[Scene.single(agent="gemini", model="gemini-3.1-flash-lite-preview")],
        environment="daytona",
    )
    result = await bf.run(config)
    print(f"Reward: {result.rewards}, Tools: {result.n_tool_calls}")

asyncio.run(main())

Note: resolve_source() is required for remote repos in the SDK. The CLI handles this transparently via --source-repo / --source-path.

API keys are auto-inherited from os.environ into the sandbox.

eval <tasks-dir> — run a benchmark suite

bench eval run \
  --source-repo benchflow-ai/skillsbench \
  --source-path tasks \
  --agent gemini \
  --model gemini-3.1-flash-lite-preview \
  --sandbox daytona \
  --concurrency 64

Or via YAML config:

bench eval run --config benchmarks/harvey-lab/harvey-lab-gemini-flash-lite.yaml

YAML format:

source:
  repo: benchflow-ai/skillsbench
  path: tasks
agent: gemini
model: gemini-3.1-flash-lite-preview
environment: daytona
concurrency: 64
max_retries: 1

metrics <jobs-dir> — analyze results

bench eval metrics jobs/      # aggregate pass-rate / tokens / cost (add --json to pipe)
bench eval list jobs/         # per-rollout table

view <rollout-dir> — view a trajectory

Results land under jobs/<job-name>/<rollout-name>/ (the default --jobs-dir is jobs/):

rollout-dir/
├── result.json              # rewards, agent, timing
├── prompts.json             # prompts sent
├── trajectory/
│   └── acp_trajectory.jsonl # tool calls + agent thoughts
└── verifier/
    ├── reward.txt           # reward value
    └── ctrf.json            # test results

create-task — create a new benchmark task

bench tasks init my-task                       # native task.md format (default)
bench tasks init my-task --no-pytest --no-oracle
bench tasks check tasks/my-task                # structural validation

Quick structure (native task.md format, the default):

my-task/
├── task.md            # YAML frontmatter (config) + prompt body
├── environment/
│   └── Dockerfile     # sandbox setup
├── verifier/
│   ├── test.sh        # verifier entrypoint -> writes /logs/verifier/reward.txt
│   └── test_outputs.py
└── oracle/            # optional reference solution (solve.sh)

--format legacy is retired in v0.6.2: bench tasks init always scaffolds a native task.md package. To bring an existing split-layout task forward, run bench tasks migrate <dir> --remove-legacy.

skills — discover and evaluate agent skills

bench skills list                                   # discover skills on disk
bench skills eval skills/citation-management \
  --agent claude-agent-acp                          # score a skill against its evals/evals.json

hub — check external-environment-hub compatibility

bench hub check          # inventory/structurally-check representative Harbor-registry tasks

agents — list available agents

bench agent list

Any agent can be prefixed with acpx/ to run via ACPX (https://acpx.sh/):

bench eval run --tasks-dir tasks/edit-pdf --agent acpx/gemini --model gemini-3.1-flash-lite-preview --sandbox daytona

ACPX is a headless ACP client with persistent sessions and crash recovery. The underlying agent's install, env vars, credentials, and skill paths are preserved.

compare — multi-agent comparison

Compare by running one config per agent (the agent: key lives in each YAML) and printing the aggregate scores:

import asyncio
from benchflow.evaluation import Evaluation

async def main():
    for config_path in [
        "benchmarks/harvey-lab/harvey-lab-gemini-flash-lite.yaml",
        "benchmarks/harvey-lab/harvey-lab-harness-parity.yaml",
    ]:
        result = await Evaluation.from_yaml(config_path).run()
        print(f"{config_path}: {result.passed}/{result.total} ({result.score:.1%})")

asyncio.run(main())

Setup

# Install benchflow from PyPI. BenchFlow CLI releases require Python 3.12+.
uv tool install --python 3.12 --upgrade benchflow
# (or from source: uv sync --extra dev --locked)
export GEMINI_API_KEY=...     # or ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.
export DAYTONA_API_KEY=...    # for cloud sandboxes

Sandboxes

Use daytona for benchmarks. Docker is limited by network exhaustion.

Skills in tasks

Two approaches for deploying skills:

Baked into Docker image (existing tasks)

COPY skills /root/.claude/skills

Runtime deployment via --skills-dir

bench eval run \
  --tasks-dir task-dir \
  --agent claude-agent-acp \
  --sandbox daytona \
  --skills-dir skills/ \
  --skill-mode with-skill

--skill-mode with-skill is required whenever you pass --skills-dir (omitting it errors). Skills are uploaded to /skills/ in the sandbox and symlinked to agent-specific paths.

Tips

• Use gemini-3.1-flash-lite-preview for testing. Use Pro/Sonnet for real benchmarks.
• Evaluations resume — re-running the same jobs_dir skips completed tasks.
• None in prompts list gets replaced with instruction.md content.
• Partial rewards work (verifier can write 0.5 to reward.txt).
• GEMINI_API_KEY requires explicit --agent-env GEMINI_API_KEY=... in CLI; SDK auto-inherits from os.environ.