Guides
All analyses › Nikolay Sekachev
You’re on YouTube for AI entrepreneurs. Read the analysis and decide whether the video is worth your time.All analyses →
How OpenAI Engineers Create Teams of 10,000 AI Agents?
Author’s title: How OpenAI Engineers Create Teams of 10,000 AI Agents?

Actually: Technical architecture for massive agent coordination using internal GPT-6 APIs currently unavailable to small businesses.

N
Nikolay Sekachev12K views · 7 days ago
5/10
BUSINESS SCORE
💭 Theory and opinions
Practicality3
Specifics7
Novelty9
✓ The analysis is enough. Summary covers the core architectural patterns for future scaling. This is a must-watch for technical leaders and developers building agentic workflows, as it provides specific architectural patterns (Sol, Astra, Luna) used by OpenAI internally. · video 24:00

🔎 What’s inside

A technical presentation from OpenAI DevDay 2026 detailing the architectural shift from simple agent loops to persistent, multi-agent systems (GPT-6) capable of massive parallel coordination for complex research.
Promise vs. reality
The title is accurate; the presentation specifically covers the evolution of agent architecture and a case study involving 10,000 agents solving the Navier-Stokes equations.

⚡ The gist in 30 seconds

  • ИИ агенты (AI agents) have evolved from simple loops requiring manual 'continue' prompts to persistent systems that manage their own goals and context compaction.
  • OpenAI uses a specialized 'Harness' and Responses API to manage tool execution and context, allowing agents to work across multiple context windows without losing their primary objective.
  • The GPT-6 architecture introduces specialized roles: Sol for core substantive work, Astra as a high-level advisor for complex reasoning, and Luna for lightweight supporting tasks.
  • Coordination patterns like 'Pair Programming' (Producer/Consumer) and 'Studio' (Department Heads) allow agents to handle long-term projects like game development autonomously.
  • Scaling to 10,000 agents requires a 'Shared Worklog' (agent.md) to maintain a unified state and prevent 'middle management' overhead between parallel agent teams.
👤 Raj Pathak (Applied AI Engineer at OpenAI) and Eric Provencher (Codex DX at OpenAI)

⏱ Key moments click — the video jumps there

Raj Pathak introduces the evolution of AI agents from simple code generation to autonomous action.
Explanation of the 'Harness' and Responses API that manages instructions and tool execution.
Deep dive into context compaction, allowing agents to summarize history and continue working when windows are full.
Introduction of the /goal command for persistent tasks that survive across multiple context resets.
Eric Provencher explains 'Ultra' mode for multi-agent collaboration on massive tasks like refactoring entire codebases.
The 'Effort Scaling' model: using Sol for core work, Astra for advice, and Luna for simple tasks.
The 'Pair Programming' pattern where a Producer agent and a Consumer agent provide mutual feedback.
The 'Studio' pattern: assigning agents as 'owners' of specific domains like Art, Design, or Programming.
Using a 'Shared Worklog' (agent.md) to keep agents on track despite frequent context compaction.
Case study: 10,000 agents working together to solve the Navier-Stokes equations.

💰 Numbers

Agents working simultaneously in Navier-Stokes research10,000 (shown on screen)
Messages exchanged between agents2.7 million (shown on screen)
Output tokens generated during the research130 billion (shown on screen)
Open math problems solved by the internal modelOver 100 (author's claim)
Numbers are quoted from the video authors.

🛠 Tools from the video

OpenAI Codex — Universal agent for coding and research tasksGPT-6 Astra — High-level advisor model for complex reasoning and reviewGPT-6 Sol — Core model for substantive work and integrationGPT-6 Luna — Lightweight model for simple supporting tasksResponses API — Internal harness for managing agent cycles, tools, and context

🚀 How to apply it in your business

  1. Adopt the 'Shared Worklog' (agent.md) pattern to help your agents maintain state in long-running automation tasks.
  2. Structure complex workflows by splitting them into 'Producer' and 'Consumer' roles to improve output quality through feedback loops.
  3. Use 'Effort Scaling' by routing simple sub-tasks to smaller, cheaper models while reserving advanced models for orchestration.
  4. Implement context compaction in your custom agent harnesses to allow them to handle tasks that exceed standard context windows.
  5. Avoid 'Middle Management' in AI architectures by keeping coordination rules simple to prevent agents from stalling on unnecessary approvals.

Who will this analysis save an hour?We’ll send the link with the “actually” takeaway — they won’t have to watch the whole video.

⚠️ Watch out: The presentation features GPT-6 and internal OpenAI APIs that may not be fully available to the public in the exact configurations shown.

❓ Questions and answers

What are ИИ агенты?

ИИ агенты (AI agents) are systems that combine reasoning with tool use to perform actions autonomously. Unlike simple chatbots, they can plan, execute tasks in a sandbox, observe results, and adjust their strategy until a goal is met.

How do you manage long-term memory in AI agents?

OpenAI uses 'context compaction' via the Responses API. When an agent's context window is full, the system summarizes the history into a compact state while keeping the primary goal and key findings, allowing the agent to continue working indefinitely.

What is the best way to coordinate multiple agents?

The video suggests several patterns: 'Manager' (one agent orchestrates sub-chats), 'Pair' (one producer, one reviewer), and 'Studio' (specialized agents owning specific domains like Art or Code). The key is to define clear boundaries and shared worklogs.

How can 10,000 agents work on one problem?

OpenAI achieved this by creating 'Agent Teams' that explore different hypotheses in parallel. These teams communicate across groups to share findings and avoid duplicating effort, all while following a unified 'Shared Worklog' to stay aligned on the final solution.

Found a mistake in the analysis? Tell us what’s wrong and we’ll fix it.
Want analyses like this for your niche? We’ll send the best videos for your business.