From ChatGPT Chatbots to Graphs: The Rapid Evolution of How We Work with LLMs

2026-07-19 · Source: AI on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, medium

Summary

Since ChatGPT's November 2022 release, the paradigm for working with large language models has rapidly evolved from conversational chatbots to sophisticated graph-based workflows. Initially, interaction focused on prompt engineering, where users crafted requests for black-box text completers. This progressed to agentic AI, enabling models to perform actions via tool use and the ReAct pattern. Production systems then adopted structured workflows for predictability, followed by context and harness engineering to manage the model's environment and runtime. A significant shift occurred around June 2026 with "loop engineering," where developers design systems that prompt agents, as highlighted by Peter Steinberger and Boris Cherny. This led to multi-agent orchestration and, currently, graph-based frameworks like LangGraph, which handle complex control flow, parallelism, and durable state. The human role has transitioned from direct prompting to designing and overseeing these advanced, goal-directed systems.

Key takeaway

For AI Engineers designing LLM-powered applications, the era of solely optimizing prompts is over. You should now prioritize designing robust "loops" or "graphs of loops" that manage agents, verify outcomes, and persist state autonomously. This shift means your focus moves to defining high-level goals, verification criteria, and accountability, allowing the system to handle inner cycles and scale complex tasks without constant manual intervention.

Key insights

LLM interaction evolved from direct prompting to designing autonomous, graph-based agent systems.

Principles

• Escalate agent autonomy only when simpler solutions fail.
• Structured workflows are a recommended starting point.
• Human role shifts to system design, goal setting, and verification.

Method

Loop engineering involves designing systems to discover work, dispatch agents, verify results, persist state, and determine next actions, replacing manual agent prompting.

In practice

• Implement ReAct pattern for agent reasoning and tool use.
• Utilize LangGraph for stateful, multi-actor graph-based workflows.
• Design outer loops with triggers, skills, and external memory.

Topics

• LLM Evolution
• Agentic AI
• Loop Engineering
• Multi-Agent Systems
• Graph-based Workflows
• LangGraph

Best for: AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• Your Hand-Written Agent Loop Won’t Survive the Real World — LangGraph structures LLM agent loops as graphs, enabling robust features like memory and human intervention.
• Environment-Grounded Automated Prompt Optimization for LLM Game Agents — The framework automatically optimizes LLM agent prompts in interactive environments using an evolutionary loop and a multi-agent decomposition.
• Another try at LLM workflow automation — Effective LLM automation requires extensive scaffolding and iterative tooling, not just smarter models.
• Agentic AI Systems in 2026: From Language Models to Autonomous Digital Workers — Agentic AI transforms LLMs into autonomous, goal-executing systems that plan, act, and adapt through iterative cognitive loops.
• The Agent Loop: How AI Goes From Answering Questions to Doing Things — Agents are LLMs in self-deciding loops, enabling autonomous task completion but introducing complexity and cost tradeoffs.
• Stop coding AI: Use Runtime Topological Self-Assembly (UC, DeepMind) — LLMs can optimize discrete symbolic graphs to discover new algorithms and dynamically architect multi-agent systems.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI on Medium.