Agent Bench

Benchmarking framework for AI coding agents on enterprise Java tasks. Defines benchmarks as YAML, launches any CLI agent, grades results with cascaded judge tiers from Agent Judge.

Quick Start

git clone https://github.com/spring-ai-community/agent-bench.git
cd agent-bench
./mvnw clean install -DskipTests

List available benchmarks:

$ bench list

Available benchmarks:

  code-coverage                  v1.0      (1 tasks)
  hello-world                    v1.0      (1 tasks)

Run a benchmark with an agent:

bench run --benchmark hello-world --agent agents/claude-code.yaml

How It Works

The bench orchestrates a per-task lifecycle:

provide → setup scripts → agent → post scripts → grade → result.json

1. Provide copies the workspace template and writes INSTRUCTION.md
2. Setup scripts run in the workspace (clone repo, compile, measure baseline)
3. Agent executes — any CLI tool that reads INSTRUCTION.md and modifies the workspace
4. Post scripts run (build, test, generate reports)
5. Grade evaluates the workspace with a cascaded jury from Agent Judge

Benchmark Format

Benchmarks live in benchmarks/ as YAML:

benchmarks/code-coverage/
├── benchmark.yaml
├── prompts/
│   └── judge-practice-adherence.txt
└── tasks/
    └── spring-petclinic/
        └── task.yaml

benchmark.yaml

Defines the jury — a cascaded sequence of judge tiers:

schema: bench.benchmark.v1
name: code-coverage
version: "1.0"
description: "Write JUnit tests to maximize JaCoCo instruction coverage."
default-timeout: PT45M

jury:
  tiers:
    - name: build
      policy: REJECT_ON_ANY_FAIL
      checks:
        - type: maven-build
          goals: [clean, test]
    - name: coverage-preservation
      policy: REJECT_ON_ANY_FAIL
      checks:
        - type: coverage-preservation
    - name: coverage-improvement
      policy: ACCEPT_ON_ALL_PASS
      checks:
        - type: coverage-improvement
          min: 50.0

task.yaml

Defines a single task — the problem, setup, and post-processing:

schema: bench.task.v1
id: spring-petclinic
instruction: |
  Write JUnit tests for this Spring Boot project to maximize code coverage.
  Run ./mvnw clean test jacoco:report to measure coverage.
  Focus on behavioral code — skip Application main classes, records, and config.
  Use narrow test slices (@WebMvcTest, @DataJpaTest) over @SpringBootTest.
timeout: PT45M
metadata:
  baselineCoverage: 0.0
setup:
  - "git init && git remote add origin https://github.com/spring-projects/spring-petclinic.git && git fetch --depth 1 origin edf4db28affc && git checkout FETCH_HEAD"
  - "./mvnw clean compile -q -Dspring-javaformat.skip=true -Dcheckstyle.skip=true"
post:
  - "./mvnw clean test jacoco:report -q -Dspring-javaformat.skip=true -Dcheckstyle.skip=true"

Agent Config

Agent configs are minimal — just a command and timeout:

# agents/claude-code.yaml
command: claude --print --dangerously-skip-permissions 'Read INSTRUCTION.md and follow the instructions precisely.'
timeout: PT45M

The bench launches the command via bash -c in the workspace directory. Any CLI tool works.

Bring Your Own Agent

The filesystem is the contract. The bench writes INSTRUCTION.md to the workspace; the agent reads it and modifies files. You can also run the provide/grade steps separately:

# Set up workspace
bench provide --benchmark code-coverage --task spring-petclinic --workspace /tmp/petclinic

# Run your agent (any tool that reads INSTRUCTION.md)
cd /tmp/petclinic && your-agent "$(cat INSTRUCTION.md)"

# Grade the result
bench grade --benchmark code-coverage --task spring-petclinic --workspace /tmp/petclinic

CLI Reference

Command	Purpose
bench list	List available benchmarks
bench tasks --benchmark <name>	List tasks in a benchmark
bench provide --benchmark <name> --task <id> --workspace <dir>	Set up workspace with instruction
bench grade --benchmark <name> --task <id> --workspace <dir>	Evaluate agent's work
bench run --benchmark <name> --agent <config> [--task <id>]	Full pipeline: provide + agent + grade

Architecture

Two modules:

• agent-bench-core — CLI, benchmark catalog, run orchestration, result model, judge factory
• agent-bench-agents — Agent-specific judge implementations (e.g., LLM-based test quality judge)

Key classes:

Class	Role
BenchmarkCatalog	Discovers benchmarks from benchmarks/ directory
BenchmarkTask	A single task: instruction, setup/post scripts, metadata
RunCommand	Orchestrates the full lifecycle per task
JudgeFactory	Materializes YAML jury config into Judge instances
TrialResult	Per-attempt result with timestamps and scores
BenchmarkResult	Aggregate result with accuracy and pass@k
ExecAgentInvoker	Loads agent config and launches the command

Module layering is enforced by ArchUnit — core does not depend on agents.

Built-in Judge Types

Type	What it checks
file-exists	File exists at path
file-content	File content matches expected (EXACT or CONTAINS)
maven-build	Maven build succeeds with specified goals
coverage-preservation	JaCoCo coverage not dropped from baseline
coverage-improvement	JaCoCo coverage exceeds threshold

Custom judge types can be registered via JudgeFactory.register().

Programmatic Usage

// Discover benchmarks
BenchmarkCatalog catalog = new BenchmarkCatalog(Path.of("benchmarks"));
List<Benchmark> benchmarks = catalog.discover();

// Find a specific benchmark
Benchmark benchmark = benchmarks.stream()
    .filter(b -> b.name().equals("code-coverage"))
    .findFirst()
    .orElseThrow();

// Access tasks
BenchmarkTask task = benchmark.tasks().get(0);
assert task.id().equals("spring-petclinic");
assert task.instruction().contains("JUnit tests");

// Wire judges from YAML config
JudgeFactory factory = new JudgeFactory();
Judge judge = factory.createFromConfig(benchmark.juryConfig());

Available Benchmarks

Benchmark	Tasks	Status
hello-world	1	Working — validates file creation
code-coverage	1 (spring-petclinic)	Judges validated, end-to-end run pending

Related Projects

• Agent Judge — Cascaded judge framework (core dependency)
• Agent Client — CLI agent integrations (Claude, Gemini)

Contributing

1. Fork the repository
2. Create a feature branch
3. Write tests for new features
4. Ensure ./mvnw clean test passes
5. Open a Pull Request

License

Apache License 2.0 — see LICENSE.