rexion-loguit 0.1.1

A minimal, explicit agent runtime

Loguit

Loguit is a Python-first, code-first runtime for building agents without LangChain's abstraction bloat. It's designed for complicated, out-there products where hidden layers get in the way.

Core principles:

• Explicit model and tool calls
• Framework-owned run state
• Run-level and step-level middleware
• Bounded state growth via pruning
• Basic structured logging

Quickstart

from loguit import Agent, init_model, tool


@tool
def add(a: int, b: int) -> int:
    return a + b


model = init_model("openai:o3-mini")
agent = Agent(model, tools=[add])

result = agent.invoke("Use add to sum 2 and 3, then answer.")
print(result.content)

agent.invoke(...) runs a single prompt through the full model→tool→model loop. agent.repl(...) launches an interactive REPL (type exit or quit to leave). Use tool_execution="parallel" when you want tool calls from a single model step to execute concurrently. Use agent.as_tool(...) to wrap a sub-agent as a normal tool.

Middleware (Run + Step)

from time import time

from loguit import BaseRunMiddleware, BaseStepMiddleware


class RunLogger(BaseRunMiddleware):
    def before_run(self, context, state):  # noqa: ANN001
        print("run start", context.run_id)


class StepTimer(BaseStepMiddleware):
    def before_step(self, context, state, step):  # noqa: ANN001
        step._started_at = time()
    def after_step(self, context, state, step, result):  # noqa: ANN001
        print(step.step_type, time() - step._started_at)

Examples

See examples/ for organized demos:

• examples/agents/
• examples/coding/
• examples/models/
• examples/mcp/
• examples/observability/
• examples/tools/

Anti-goals

Loguit is intentionally narrow in V1:

• No LangChain API compatibility
• No graph DSLs or chain abstractions
• No RAG or vector store framework
• No durable state persistence
• No multi-agent orchestration

Why Loguit?

Because you should be able to build complex products without being chained to bloatware.