Agent Systems · R&D Demos

Multi-Agent Content Orchestration

Exploring deterministic agent handoffs and human-in-the-loop control with LangGraph

Juan Iturbe •
Exploring deterministic agent handoffs and human-in-the-loop control with LangGraph

Status

Research & Demo

This is an experimental LangGraph implementation built to explore agent orchestration patterns such as stateful workflows, deterministic handoffs, evaluation gates, and human-in-the-loop control.

Overview

ReAct-swarm is a stateful multi-agent pipeline where specialized agents collaborate through explicit state transitions rather than ad-hoc prompt chaining.

The goal of the project is to answer a practical question:

What does an agent system need once you move past single-prompt demos?

The system focuses on control, predictability, and recovery—the parts that tend to break first in real agent workflows.

Multi Agent Architecture

Architecture Summary

The workflow is orchestrated using LangGraph as a deterministic state machine.

Agents:

  • Draft – produces an initial draft from user input and source data
  • Refiner – improves clarity and structure while preserving intent
  • Editor – evaluates output against a fixed quality rubric

Key constraints:

  • Only one agent can act at a time
  • All handoffs are explicit
  • Human approval is required at defined checkpoints
  • Evaluation decisions are deterministic and auditable

Cognitive Architecture

Workflow

Draft
  → (human approval)
Refiner
  → (human approval)
Editor (rubric-based evaluation)
  → approve → publish
  → reject → back to Draft

The graph allows controlled loops (Editor → Draft) without losing state.

Main Characteristics

Stateful Agent Orchestration

Agents operate within a shared state model rather than isolated calls.

Deterministic Handoffs

Each transition is intentional and traceable. There are no implicit agent jumps or hidden autonomy.

Human-in-the-Loop Checkpoints

Approval gates allow intervention without breaking the workflow or resetting context.

Evaluation as a Quality Gate

The Editor agent scores output against a fixed 7-point rubric. Clear thresholds determine whether content advances or loops back.

Persistent State & Recovery

Agent state is checkpointed in PostgreSQL, enabling:

  • Recovery from failures
  • Auditability
  • Replay of past runs

Technology Stack

  • LangGraph — stateful agent orchestration
  • OpenAI GPT-4 — generation and evaluation
  • PostgreSQL — state persistence and recovery
  • TypeScript — type-safe agent contracts
  • LinkedIn API — example external integration (non-core)

Explore the Code

  • GitHub: https://github.com/ituPhi/ReAct-swarm

  • Relevant areas:

    • src/graph.ts — orchestration logic
    • src/state.ts — shared state model
    • src/agents/ — agent definitions
    • db/ — persistence and configuration
Ver código