Introduction

What is Coevolved?

Coevolved is a lightweight Python framework for building multi-step, multi-agent runtimes with:

• A single core primitive: Step (an atomic unit of computation)
• Composable execution: sequence, parallel, fallback, retry
• First-class observability: structured trace events with pluggable sinks
• Production controls: budgets/timeouts, checkpointing, interrupts (HITL)
• Agent patterns: reusable loop infrastructure + a prebuilt ReAct agent

How Coevolved is structured

The Python package is split into three layers:

• coevolved.base: foundational runtime primitives (Step, tracing, composition, checkpointing, interrupts, policies)
• coevolved.core: LLM + tool integration (llm_step, tool_step, prompt utilities, provider interfaces)
• coevolved.prebuilt: ready-to-use agent patterns (loop + ReAct) built from base/core primitives

Key concepts

Steps

Everything is a Step: pure functions, tools, LLM calls, and even agents (as loops) are represented as steps.

State + schemas

Steps operate on a shared “state” (often a dict). You can optionally validate inputs/outputs with Pydantic models.

Composition

Build workflows by composing steps (e.g. sequential pipelines, retries, or fallback chains) instead of writing a bespoke runtime each time.

Tracing

Each step execution emits lifecycle events (start, end, error) and can capture snapshots (typically on error) for debugging.

Agents as loops

An “agent” is usually a loop that repeatedly runs an iteration body until a stop condition is met—optionally with checkpointing and budget enforcement.

Next steps

Quickstart

Build your first workflow with Steps, composition, and an agent loop.

Steps

Learn the single core primitive that powers tools, LLM calls, and agents.