Harbor Agents

A collection of custom agents for Harbor - the AI agent evaluation framework.

Installation

pip install harbor-agents

Available Agents

ClaudeCodeWithSkills

Extends Claude Code with custom skills support. Load pre-configured skills into the container for evaluations.

# Load all skills from a directory
harbor run -p ./my-task \
    --agent-import-path harbor_agent.skilled_claude:ClaudeCodeWithSkills \
    -m anthropic/claude-sonnet-4-20250514 \
    --ak skill_dir=./skills

# Load specific skills only
harbor run -p ./my-task \
    --agent-import-path harbor_agent.skilled_claude:ClaudeCodeWithSkills \
    -m anthropic/claude-sonnet-4-20250514 \
    --ak skill_dir=./skills \
    --ak skills=my-skill,another-skill

# Baseline (no skills)
harbor run -p ./my-task \
    --agent-import-path harbor_agent.skilled_claude:ClaudeCodeWithSkills \
    -m anthropic/claude-sonnet-4-20250514 \
    --ak skill_dir=./skills \
    --ak 'skills='

Options:

Skill Directory Structure:

skills/
├── my-skill/
│   ├── SKILL.md          # Required
│   └── references/       # Optional
└── another-skill/
    └── SKILL.md

MultiTurnAgent

A composite agent for testing multi-turn conversations. Combines a simulated user (which generates prompts) with an inner agent (which processes them).

harbor run -p ./my-task \
    --agent-import-path harbor_agent.multi_turn:MultiTurnAgent \
    -m anthropic/claude-sonnet-4-20250514 \
    --ak simulated_user=my_module:MySimulatedUser \
    --ak agent=harbor.agents.installed.claude_code:ClaudeCode \
    --ak 'agent_kwargs={"model_name": "anthropic/claude-sonnet-4-20250514"}' \
    --ak max_turns=10

Options:

Creating a Simulated User:

from harbor_agent.multi_turn import SimulatedUser, SimulatedUserDone, ConversationMessage

class MySimulatedUser(SimulatedUser):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self._turn = 0

    async def next_message(self, conversation: list[ConversationMessage]) -> str:
        self._turn += 1
        if self._turn > 3:
            raise SimulatedUserDone("Task complete")
        return f"Please do step {self._turn}"

Conversation Flow:

1. next_message() is called with conversation history
2. Returns a prompt string → sent to inner agent
3. Inner agent responds → added to history
4. Repeat until SimulatedUserDone is raised or max_turns reached

Output: Saves trajectory.json in ATIF format with all conversation turns.

Development

uv sync --all-extras    # Install
uv run pytest -v        # Test
uv run ruff check src   # Lint
uv run mypy src         # Type check

License

Apache 2.0