Tech • AI • Robotics • Game
What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026
9.1/10TL;DR
OpenAI outlined how modern coding agents evolved from simple tool use into persistent, multi-agent systems that can work across long tasks, coordinate subagents, and even tackle research problems at the scale of 10,000 agents.
KEY POINTS
From chat outputs to action-taking agents
Early language models could generate working code for simple tasks such as FizzBuzz, but users still had to copy, paste, run, and debug the output themselves. The major shift came when models combined reasoning with tool use, allowing them to inspect results, adapt plans, and take the next step without human intervention.
Codex built around a model-plus-harness design
Codex was designed as a general computer-using agent rather than a narrow workflow tool. Its core architecture pairs the model with a harness that manages instructions, tool definitions, execution in a sandbox, conversation state, and decisions about whether the agent should continue working or stop.
Minimal tools, broad generalization
The design philosophy emphasized a small set of general tools such as the shell, file system access, and computer controls. The aim was to let the model solve tasks much as a human would on a computer, so that if a person could complete a task in that environment, the agent could learn to figure it out too.
The Responses API powers the loop
At the center of the system is the Responses API, which runs the agentic loop. The model requests tool calls, the harness executes them in the sandbox, appends the results back into context, and decides whether to trigger another step, creating a managed cycle of reasoning, action, and feedback.
Compaction enabled long-horizon work
One of the biggest engineering challenges was context length. As agents read files, run tests, and accumulate history, the context window fills up, so compaction was introduced to condense prior work into a summary state while preserving what matters for future steps. Codex was trained to understand compacted states and resume effectively after them, enabling work across multiple context windows.
Persistence replaced brute-force looping
Earlier systems often needed external prompts telling them to continue, a pattern widely known as loop engineering. OpenAI first addressed this with dash goal, which stored the objective outside the main conversation so it would survive compaction and repeated turns, then later trained persistence directly into the model. By GPT-5.6, the model could keep going, check its own progress, and stop only when blocked or finished.
Single agents first, multi-agent systems only when needed
For many tasks, a single strong agent is enough if it has a clear objective and a way to verify its work. More agents are useful mainly for decomposable workloads such as reviewing several pull requests or splitting a broad investigation, but each added agent can sharply increase token usage without guaranteeing better results.
GPT-6 expanded coordination between agents
With GPT-6, agents were trained to coordinate and collaborate, breaking large problems into independent pieces and multitasking across them. In Ultra configurations, Codex can create copies of itself, assign subtasks, and maintain separate working contexts, though users are still encouraged to be deliberate because unnecessary hierarchies can create a costly middle management tax.
Thread managers and work logs improve control
Beyond subagents, Codex can manage multiple ongoing chat threads, steering separate workstreams while keeping the overall effort synchronized. A compact work log can also preserve the main objective, current progress, and user steering across repeated compaction events, reducing drift during long-running projects such as games or large refactors.
Large-scale agent teams are now being used for research
OpenAI described an internal effort on the Navier-Stokes problem that used 10,000 agents, exchanged 2.7 million messages, and generated about 130 billion output tokens. The architecture grouped agents into teams pursuing different approaches under one shared objective, with cross-team communication used to share findings, avoid duplication, and redirect effort toward the most promising lines of inquiry.
Results extend beyond coding workflows
In the Navier-Stokes effort, teams explored hypotheses, challenged assumptions, ran experiments, and spawned new teams as new questions emerged. OpenAI said the same internal model behind that work has helped resolve more than 100 longstanding problems across mathematics and continues to be used in math and science discovery.
CONCLUSION
The direction of travel is from scripted loops toward autonomous systems that can persist, coordinate, and scale their effort. The main challenge is no longer whether agents can act, but how to structure that action efficiently as tasks grow longer, broader, and more collaborative.
Ask a question