wbal 0.3.2

Weights & Biases Agent Library - A minimal framework for building LLM agents

WBAL

Weights & Biases Agent Library - A minimal framework for building LLM agents.

Overview

WBAL provides three core primitives:

• Agent - Orchestrates the perceive-invoke-do loop
• Environment - Provides tools and context
• LM - Language model interface

All components inherit from WBALObject (Pydantic BaseModel + observe() method).

Quick Start

Requires OPENAI_API_KEY in your environment.

import weave
from wbal import Environment, OpenAIWBAgent, weaveTool, GPT5MiniTester

weave.init('my-project')

class MyEnv(Environment):
    env = "You are a helpful assistant."
    include_tools_in_observe = True

    @weaveTool
    def greet(self, name: str) -> str:
        """Greet someone by name."""
        return f"Hello, {name}!"

agent = OpenAIWBAgent(
    lm=GPT5MiniTester(),
    env=MyEnv(task="Say hello to Alice"),
    maxSteps=5,
    system_prompt="Use tools when helpful. Call exit() when you're done.",
)
agent.run()

Installation

From PyPI:

pip install wbal

From source (for local development):

git clone <this-repo>
cd wbal
uv sync

Documentation

Document	Description
USER.md	Usage guide, API reference, examples
DEVELOPER.md	Architecture, contributing, testing
Agent_Instructions.md	Agent/environment guidance

CLI

# Run (baseline, non-interactive)
wbal run --project my-project --task "Say hello to Alice, then call exit()"

# Chat (interactive via tool calls)
wbal chat --project my-project --task "Say hello to Alice"

# Poll (runs once or on an interval)
wbal poll --project my-project --task "Check status" --interval 300

# Run from a YAML agent manifest
wbal run --project my-project --agent-spec path/to/agent.yaml --task "Do the thing"

YAML Agents

WBAL supports YAML-based agent manifests that let you configure:

• model + max steps
• prompt files (YAML)
• tool modules to attach to the agent/env
• explicit subagent delegation (DAG) via run_agent

See examples/agents/README.md.

Agent Bundles (WandBSwarm-compatible)

WBAL can run “agent bundles” that expose run.sh (required) and install.sh (optional), with the same env var contract as WandBSwarm (AGENT_DIR, TASK_DIR, WORKSPACE, etc.).

wbal bundle validate --agent-dir path/to/agent
wbal bundle run --agent-dir path/to/agent --task-dir path/to/task --workspace-dir ./workspace

API

from wbal import (
    # Core
    Agent, Environment, StatefulEnvironment, LM,

    # Models
    GPT5Large, GPT5MiniTester,

    # Decorators
    weaveTool, tool,

    # Mixins
    ExitableAgent,

    # Helpers
    tool_timeout, format_openai_tool_response,
)

Examples

See examples/ for complete implementations.

• examples/zagent_v1.py - Orchestrator-style agent with persistent notes + bash tool

Run locally:

uv run python examples/zagent_v1.py --task "Inspect this repo, take notes, then exit()"

Structure

wbal/
├── wbal/
│   ├── agent.py        # Agent class
│   ├── environment.py  # Environment, StatefulEnvironment
│   ├── lm.py           # LM, GPT5Large, GPT5MiniTester
│   ├── helper.py       # Tool decorators and utilities
│   └── mixins.py       # ExitableAgent
├── tests/
└── examples/