# What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026

**URL:** <https://www.artofsm.art/t/what-comes-after-loops-deep-dive-into-agentic-architecture-devday-2026/28124>\
**Category:** Content Creators\
**Tags:** codex, agents, open-ai, machine-learning, automation, openai\
**Created:** [11 November 2026 01:24 UTC](https://www.artofsm.art/t/what-comes-after-loops-deep-dive-into-agentic-architecture-devday-2026/28124 "2026-11-11T01:24:07Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\
**Post date:** [11 November 2026 01:24 UTC](https://www.artofsm.art/t/what-comes-after-loops-deep-dive-into-agentic-architecture-devday-2026/28124/1 "2026-11-11T01:24:07Z")

</div>

[![](https://www.artofsm.art/uploads/default/original/3X/0/8/08bc411c79113421f0abf166dc6245d44e06b565.jpeg "What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026") ](https://www.youtube.com/watch?v=nAubbmlhPfI)

In “What Comes After Loops?: Deep Dive Into Agentic Architecture,” Raj Bhatek and Eric from OpenAI explore the evolution of AI agents from simple code generation to sophisticated, persistent multi-agent systems capable of complex problem-solving and large-scale collaboration. They highlight advancements like Codex’s autonomous coding loop, persistence training in GPT-5.6, and the coordination of thousands of agents on scientific challenges, emphasizing the transformative potential of agentic architectures for future AI workflows.

---

<div class="post-metadata">

**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\
**Post date:** [11 November 2026 01:44 UTC](https://www.artofsm.art/t/what-comes-after-loops-deep-dive-into-agentic-architecture-devday-2026/28124/3 "2026-11-11T01:44:43Z")

</div>

In the presentation “What Comes After Loops?: Deep Dive Into Agentic Architecture,” Raj Bhatek and Eric from OpenAI explore the evolution of AI agents from simple code generation to sophisticated multi-agent systems capable of complex problem-solving. Raj begins by recounting the early days of ChatGPT, highlighting the initial excitement when AI could generate working code like FizzBuzz, but noting the limitation that users still had to manually transfer code to their environments. The real breakthrough came with Codex, an AI agent designed to act autonomously by reasoning, using tools, and interacting with a computer environment through a tightly integrated harness that manages instructions, tools, and context. This agentic loop allowed Codex to write, run, and debug code independently, marking a significant step forward in AI capabilities.

Raj further explains the importance of managing context and persistence in AI agents. Since models have limited context windows, OpenAI developed a compaction mechanism that condenses conversation history without losing essential information, enabling agents to work over long horizons and multiple compaction events. Early models struggled with task completion, often ending prematurely, which led to the development of loop engineering techniques like the Ralph loop and the dash goal system. These methods ensured agents could persistently work on tasks until completion. Eventually, persistence was trained directly into GPT-5.6, allowing the model to autonomously continue working without external nudges, significantly enhancing its reliability and effectiveness.

Eric then takes over to discuss practical applications and best practices for using Codex today. He advises starting with a single agent for clarity and efficiency before scaling to multiple agents when tasks demand it. With GPT-6, agents can coordinate and collaborate by decomposing complex problems into manageable parts and multitasking through self-replication and communication. However, Eric cautions that adding more agents increases token usage and complexity, so it’s crucial to balance the benefits against the costs. He shares examples of agent structures, such as pairs working on interdependent tasks and thread managers that oversee multiple concurrent workstreams, emphasizing the importance of clear communication and avoiding over-management to prevent inefficiencies.

The discussion then shifts to large-scale agent teams, exemplified by OpenAI’s work on the Navier-Stokes problem, where 10,000 agents collaborated through 2.7 million messages and 130 billion tokens to solve a longstanding mathematical challenge. This scale required sophisticated coordination, with teams working on different hypotheses, experiments, and approaches while maintaining communication to avoid duplicated efforts. The architecture allowed dynamic reallocation of resources and focus as new insights emerged, demonstrating how agent teams can tackle ambitious scientific and mathematical problems by combining deep and broad research efforts under a unified goal.

In closing, Raj challenges the audience to apply these advanced agentic architectures to problems they have struggled with, encouraging experimentation with multi-agent collaboration and persistence techniques. He highlights that while best practices evolve rapidly, the potential for AI agents to transform complex workflows and discovery processes is immense. The presentation underscores that the journey from simple loops to large-scale agent teams marks just the beginning of what AI can achieve, inviting developers and researchers to build on these foundations and push the boundaries of autonomous AI systems.

## Useful Links

- [OpenAI Codex GitHub Repository](https://github.com/openai/codex) — Directly substantiates and enables action on the design and use of Codex agents.
- [Research on Agentic Architectures and Loop Engineering by OpenAI](https://www.ibm.com/think/topics/loop-engineering) — Provides foundational research and explanations on the agentic loop and persistence mechanisms discussed.
- [OpenAI GPT-6 Model Announcement and Technical Details](https://developers.openai.com/api/docs/models) — Details the model capabilities that underpin the multi-agent systems and persistence features discussed.
- [OpenAI Navier-Stokes Problem Solution Using Multi-Agent Systems](https://openai.com/index/navier-stokes-solution/) — Describes the large-scale agent team architecture and coordination used to solve a complex scientific problem.
