Plexus Procedures

Configuration-based agentic workflow programming with first-class human-in-the-loop support.

What are Procedures?

Plexus Procedures is a domain-specific language (DSL) that enables you to define sophisticated agentic workflows through configuration, combining:

  • Declarative YAML for component definitions (agents, prompts, tools, stages)
  • Embedded Lua for orchestration logic and control flow
  • High-level primitives that abstract LLM mechanics (e.g., Worker.turn())
  • First-class HITL (Human-in-the-Loop) support for collaboration

Key Features

Uniform Recursion

Procedures work identically at all nesting levels - same parameters, outputs, HITL, and async capabilities everywhere.

Human-in-the-Loop

First-class primitives for approval, input, review, notifications, and alerts built directly into workflow control flow.

Async Operations

Spawn procedures in parallel with simple primitives. Monitor progress and wait for results with built-in concurrency support.

Built-In Reliability

Automatic retries, validation, error handling, and recovery patterns under the hood.

Quick Example

Here's a simple procedure that researches a topic and requests human approval:

name: content_reviewer
version: 1.0.0

params:
  topic:
    type: string
    required: true

outputs:
  approved:
    type: boolean
    required: true

agents:
  researcher:
    system_prompt: |
      Research and summarize: {params.topic}
    tools: [search, done]

workflow: |
  -- AI does research
  repeat
    Researcher.turn()
  until Tool.called("done")

  -- Human approves
  local approved = Human.approve({
    message = "Approve this research?",
    context = {findings = State.get("findings")}
  })

  return {approved = approved}

Core Concepts

Procedures

Reusable units of agentic work that accept typed parameters, return typed outputs, and can invoke other procedures recursively.

Agents

Configured LLM instances with system prompts, tools, and behavior settings. Each agent becomes a Lua primitive (e.g., Worker.turn()).

Workflows

Lua orchestration code that controls execution flow, manages state, invokes agents, and handles human interaction.

State & Stages

Mutable storage for tracking progress across turns, and high-level workflow phases for monitoring and status reporting.

Message Classification

Every message in Plexus has a humanInteraction classification that determines visibility and behavior:

  • INTERNAL - Agent reasoning hidden from humans
  • CHAT / CHAT_ASSISTANT - Conversational AI messages
  • NOTIFICATION - Workflow progress updates
  • ALERT_* - System monitoring alerts
  • PENDING_* - Requests requiring human response

This unified system enables different UIs to filter messages appropriately - operator dashboards see notifications and requests, monitoring sees alerts, debugging sees everything.

Documentation

Getting Started

Learn the basics and write your first procedure

Human-in-the-Loop

Master collaboration patterns with humans

Examples & Patterns

Real-world examples from simple to complex

API Reference

Complete reference for all Lua primitives

Message Classification

Understand message types and visibility

Technical Specification

Complete DSL specification and reference

Additional Resources

AGENTS.md: Comprehensive technical specification located at /plexus/procedures/AGENTS.md

HTML Documentation: Complete standalone HTML documentation available at /plexus/procedures/docs/

GitHub: github.com/AnthusAI/Plexus