Agent configuration - CooperBench

The --agent-config flag allows you to provide agent-specific configuration files to customize agent behavior.

Usage

cooperbench run --agent-config <path-to-config>

Configuration format

Configuration format depends on the agent framework:

mini_swe_agent - YAML format
swe_agent - YAML format
openhands - YAML or TOML format

mini_swe_agent configuration

Example config

mini.yaml

agent:
  system_template: |
    You are a helpful assistant that can interact with a computer.
    {% if agent_id %}
    You are {{ agent_id }} working as a team with: {{ agents | join(', ') }}.
    {% endif %}
    
    Your response must contain exactly ONE bash code block with ONE command.
  
  instance_template: |
    Please solve this issue: {{task}}
    
    You can execute bash commands and edit files.
  
  step_limit: 100
  cost_limit: 1.0
  mode: confirm

environment:
  env:
    PAGER: cat
    MANPAGER: cat

Key fields

system_template - System prompt for the agent (supports Jinja2 templates)
instance_template - Task-specific instructions (supports Jinja2 templates)
step_limit - Maximum number of agent turns (default: 100)
cost_limit - Maximum cost in USD (default: 1.0)
mode - Execution mode (confirm, auto)

Template variables

Available in Jinja2 templates:

{{task}} - Task description
{{agent_id}} - Agent identifier (e.g., agent_0)
{{agents}} - List of all agent IDs
{{git_enabled}} - Whether git collaboration is enabled
{{messaging_enabled}} - Whether messaging is enabled
{{system}}, {{release}}, {{version}}, {{machine}} - System info

Example usage

cooperbench run \
  --agent mini_swe_agent \
  --agent-config configs/mini.yaml

swe_agent configuration

Example config

coop.yaml

agent:
  templates:
    system_template: |-
      You are a helpful assistant that can interact with a computer to solve tasks.
      {% if agent_id and agents|length > 1 %}
      You are {{ agent_id }} working as a team with: {{ agents | join(', ') }}.
      {% endif %}
    
    instance_template: |-
      I've uploaded a python code repository in the directory {{working_dir}}.
      
      <pr_description>
      {{problem_statement}}
      </pr_description>
      
      Please implement the necessary changes.
  
  tools:
    env_variables:
      PAGER: cat
      MANPAGER: cat
    bundles:
      - path: tools/registry
      - path: tools/edit_anthropic
      - path: tools/messaging
    enable_bash_tool: true
  
  history_processors:
    - type: cache_control
      last_n_messages: 2

Key fields

templates - Prompt templates (system, instance, observation)
tools - Tool configuration (bundles, env variables)
history_processors - Message history processing

Tool bundles

Available tool bundles:

tools/registry - Core tools (ls, find, grep, etc.)
tools/edit_anthropic - File editing tools
tools/messaging - send_message command for collaboration
tools/review_on_submit_m - Code review on submission

Example usage

cooperbench run \
  --agent swe_agent \
  --agent-config configs/coop.yaml \
  --setting coop

openhands configuration

OpenHands uses YAML or TOML configuration.

Example config

openhands.yaml

agent:
  name: CodeActAgent
  max_iterations: 100
  enable_tools:
    - terminal
    - editor
    - browser

llm:
  model: gpt-4o
  temperature: 0.1
  max_tokens: 4096

security:
  enable_sandbox: true
  allow_network: true

Example usage

cooperbench run \
  --agent openhands \
  --agent-config configs/openhands.yaml

Environment-specific configs

Solo mode config

For single-agent tasks, simplify templates:

solo.yaml

agent:
  system_template: |
    You are a helpful coding assistant.
    Solve the task by executing bash commands and editing files.
  
  instance_template: |
    Task: {{task}}
    
    Workflow:
    1. Analyze the codebase
    2. Reproduce the issue
    3. Fix the issue
    4. Verify the fix
    5. Submit: `echo COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT`
  
  step_limit: 50

Usage:

cooperbench run --setting solo --agent-config configs/solo.yaml

Cooperative mode config

For multi-agent tasks, include collaboration instructions:

coop.yaml

agent:
  system_template: |
    You are {{ agent_id }} working with: {{ agents | join(', ') }}.
    Each agent has their own workspace.
    {% if messaging_enabled %}
    Use send_message <agent_name> "message" to coordinate.
    {% endif %}
    {% if git_enabled %}
    Use shared git remote "team" to share code.
    {% endif %}
  
  instance_template: |
    Task: {{task}}
    
    Collaboration workflow:
    1. Explore codebase
    2. Message teammates about which files you'll modify
    3. Implement your feature
    4. Push changes and submit

Usage:

cooperbench run --setting coop --agent-config configs/coop.yaml --git

Default configurations

If --agent-config is not specified, CooperBench uses built-in defaults:

mini_swe_agent: src/cooperbench/agents/mini_swe_agent/config/mini.yaml
swe_agent: src/cooperbench/agents/swe_agent/config/coop.yaml
openhands: Built-in SDK defaults

Locating default configs

To see default configs:

# mini_swe_agent default
cat ~/.local/share/uv/cooperbench/agents/mini_swe_agent/config/mini.yaml

# swe_agent default
cat ~/.local/share/uv/cooperbench/agents/swe_agent/config/coop.yaml

Or from source:

cat src/cooperbench/agents/mini_swe_agent/config/mini.yaml
cat src/cooperbench/agents/swe_agent/config/coop.yaml

Custom config examples

Increase step limit

high-steps.yaml

agent:
  system_template: |
    You are a helpful coding assistant.
  instance_template: |
    Task: {{task}}
  step_limit: 200  # Default is 100
  cost_limit: 2.0  # Default is 1.0

cooperbench run --agent-config high-steps.yaml

Custom prompts

custom-prompt.yaml

agent:
  system_template: |
    You are an expert Python developer.
    Write clean, well-documented code.
    Always add type hints and docstrings.
  
  instance_template: |
    Task: {{task}}
    
    Requirements:
    - Add comprehensive docstrings
    - Include type hints
    - Write unit tests
    - Follow PEP 8 style

cooperbench run --agent-config custom-prompt.yaml

Disable tools

minimal-tools.yaml

agent:
  tools:
    env_variables:
      PAGER: cat
    bundles:
      - path: tools/registry  # Core tools only
    enable_bash_tool: true

cooperbench run --agent swe_agent --agent-config minimal-tools.yaml

Environment variable overrides

Some settings can be overridden with environment variables:

# Override config directory
export MSWEA_CONFIG_DIR=./configs
cooperbench run --agent-config mini.yaml

See Environment variables for full list.

Validation

CooperBench validates configs on load. Common errors: Invalid YAML syntax:

Error: Failed to parse config: mapping values are not allowed here

Fix: Check indentation and syntax. Missing required fields:

Error: Missing required field: agent.system_template

Fix: Add the required field to your config. Invalid template:

Error: Jinja2 template error: 'task' is undefined

Fix: Check template variable names match available variables.

Best practices

Start with defaults: Copy built-in configs and modify incrementally
Test locally: Use --backend docker -c 1 to test configs quickly
Version control: Keep configs in git alongside your experiments
Document changes: Add comments explaining custom settings
Validate templates: Test Jinja2 templates with sample data before running

Commands

Configuration

​Usage

​Configuration format

​mini_swe_agent configuration

​Example config

​Key fields

​Template variables

​Example usage

​swe_agent configuration

​Example config

​Key fields

​Tool bundles

​Example usage

​openhands configuration

​Example config

​Example usage

​Environment-specific configs

​Solo mode config

​Cooperative mode config

​Default configurations

​Locating default configs

​Custom config examples

​Increase step limit

​Custom prompts

​Disable tools

​Environment variable overrides

​Validation

​Best practices

Usage

Configuration format

mini_swe_agent configuration

Example config

Key fields

Template variables

Example usage

swe_agent configuration

Example config

Key fields

Tool bundles

Example usage

openhands configuration

Example config

Example usage

Environment-specific configs

Solo mode config

Cooperative mode config

Default configurations

Locating default configs

Custom config examples

Increase step limit

Custom prompts

Disable tools

Environment variable overrides

Validation

Best practices