abstractagent 0.3.0

Agent implementations using AbstractRuntime and AbstractCore

AbstractAgent

Agent implementations using AbstractRuntime and AbstractCore.

Features

• ReAct Agent: Reason-Act-Observe loop with tool calling
• Host UX in AbstractCode: the interactive terminal shell lives in AbstractCode; AbstractAgent stays focused on agent patterns/workflows
• Pause/Resume: Durable agent state with interrupt/resume capability
• Ask User: Agent can ask questions with multiple choice + free text
• Ledger Recording: All tool calls recorded for auditability

Installation

pip install -e .

Quick Start

Simple (Factory)

from abstractagent import create_react_agent

# One-liner agent creation
agent = create_react_agent()
agent.start("List the files in the current directory")
state = agent.run_to_completion()
print(state.output["answer"])

With Custom Tools

from abstractagent import create_react_agent
from abstractcore.tools import tool

@tool(name="my_tool", description="My custom tool")
def my_tool(query: str) -> str:
    """My custom tool."""
    return f"Result for {query}"

agent = create_react_agent(tools=[my_tool])

Full Control

from abstractruntime.integrations.abstractcore import create_local_runtime
from abstractagent import ReactAgent, list_files, read_file

# Create runtime
runtime = create_local_runtime(
    provider="ollama",
    model="qwen3:4b-instruct-2507-q4_K_M",
)

# Create agent with specific tools
agent = ReactAgent(
    runtime=runtime,
    tools=[list_files, read_file],
)

agent.start("List the files in the current directory")
state = agent.run_to_completion()
print(state.output["answer"])

State Persistence

Resume agents across process restarts:

agent = create_react_agent()
agent.start("Long running task")

# Save state before exit
agent.save_state("agent_state.json")

# ... process restarts ...

# Load and resume
agent = create_react_agent()
agent.load_state("agent_state.json")
state = agent.run_to_completion()

# Cleanup
agent.clear_state("agent_state.json")

Interactive Shell (AbstractCode)

# The interactive REPL moved to AbstractCode (host UX).
abstractcode --agent react --provider ollama --model qwen3:4b-instruct-2507-q4_K_M

Architecture

AbstractAgent
     │
     ├── Uses AbstractRuntime for durable execution
     │   - Workflows survive crashes
     │   - Pause/resume capability
     │   - Ledger tracks all actions (LLM calls, tool calls)
     │
     └── Uses AbstractCore for LLM/tools
         - Provider-agnostic LLM calls
         - Tool registration and execution
         - Tool call parsing for all model architectures

Available Tools

• Default tool callables are re-exported from AbstractCore in abstractagent.tools (file ops, web tools, execute_command), plus:
  • execute_python(code, timeout_s=...)
  • self_improve(suggestion, ...)
• The agent also exposes schema-only built-ins (ask_user, recall_memory, inspect_vars, remember, compact_memory) which are translated into Runtime effects by the workflow adapters (durable; no callable persistence).