AgentCore is the reasoning engine at the heart of CLIver. It’s a self-contained Python class that implements the Re-Act loop — the cycle of sending messages to an LLM, executing tools, and continuing until a final response is ready.

Why It Matters

AgentCore has zero CLI dependencies. It takes config dicts and returns message objects. This means you can embed it directly into your own Python applications, web services, or automation scripts.

The Re-Act Loop

1. Receive user input
2. Build messages (system prompt + conversation history + user input)
3. Send to LLM
4. If LLM returns a tool call → execute tool → go to step 3
5. If LLM returns text → done, return response

The loop runs up to 50 iterations by default, with a safety limit of 5 consecutive tool errors before stopping.

Using AgentCore in Python

from cliver.agent_factory import create_agent_core
from cliver.config import ConfigManager

# Load config (auto-loads from ~/.cliver)
config = ConfigManager()
model = config.get_llm_model()

# Create agent — tools, MCP, and identity are derived automatically
agent = create_agent_core(model_config=model)

# Chat (non-streaming)
response = await agent.chat("What's the weather in Tokyo?")
print(response.message.text)

# Stream (real-time output)
async for chunk in agent.stream("Write a poem about code"):
    if chunk.content:
        print(chunk.content, end="", flush=True)

Key Properties

• Minimal construction. Just a ModelConfig — tools, MCP client, and agent identity are derived from the config directory automatically.
• Stateless instance, stateful calls. One AgentCore instance per model. Conversation context is passed per call — safe for concurrent use.
• Tool gathering is lazy. Builtin tools are registered at construction. Extra tools, MCP servers, and tool filtering are resolved at call time via chat()/stream().
• Events are observable. Pass an on_event handler to monitor tool calls, errors, and inference in real time.
• Media generation. Call agent.generate(prompt, media_type="image") to produce images, audio, or video via the provider’s generation endpoint.