How Third Moment Research Runs 10,000 AI Agents Simultaneously for Over 36 Hours

While most AI applications still struggle to maintain conversational coherence for just a few minutes, Third Moment Research has pushed agent runtime beyond 36 hours — and not just one or two agents, but 10,000 running simultaneously.

This AI research lab's approach reveals the critical path for agents transitioning from "toys" to "productivity tools."

Core Capability: Making Agents "Focus for Days"

The defining feature of Third Moment Research's agent system isn't the quality of any single task completion — it's the stability and focus of continuous operation. According to their public information, these agents can "focus for days" — maintaining task focus at a multi-day scale — and possess the ability to "improve with use," continuously optimizing their own performance during operation.

This combination of capabilities addresses a key industry pain point: most current AI agents are prone to goal drift, context loss, or performance degradation during long-horizon tasks. Third Moment Research's approach demonstrates that through sound architectural design, agents can maintain effective operation over timescales of tens of hours.

Running 10,000 agents simultaneously for over 36 hours means the system must handle massive concurrent state management, resource scheduling, and consistency assurance challenges. This isn't simple horizontal scaling — it's a comprehensive test of the underlying infrastructure and agent orchestration framework.

Goal drift and context loss are the most typical failure modes in long-horizon agent systems. Goal drift refers to agents gradually deviating from their initial objectives during multi-step task execution — for example, a monitoring agent originally tasked with tracking anomalous transactions might, hours later, start optimizing log formats instead. Context loss stems from the limited context windows of large language models: when runtime exceeds window capacity, early critical information gets pushed out, causing agents to "forget" task constraints. Common mitigation approaches include external memory storage (such as vector databases), periodic goal re-injection, and hierarchical task decomposition. Third Moment Research's claimed "focus for days" capability fundamentally requires systematically solving both categories of problems at the architectural level, rather than relying solely on model capability improvements.

Technical Foundation: Deep Integration of the GPT-6 Model Family

Third Moment Research's tech stack is built around GPT-6 Astra and GPT-6.1 Sol, while also integrating two specialized systems: Codex and Daybreak.

Codex handles code-related task processing, while Daybreak focuses on the cybersecurity domain. This "general base + vertical specialization" architecture allows agents to handle open-ended tasks while achieving professional-level performance in specific domains.

Notably, the progression between model versions — the coexistence of Astra and Sol — suggests a layered strategy: tasks of different complexity may be handled by models of different capabilities, achieving a balance between cost and effectiveness.

The "general base + vertical specialization" architectural pattern is becoming increasingly common in AI engineering. The core idea is: use the most capable general-purpose large model for open-ended reasoning and planning tasks, while equipping specific domains (such as code generation or security analysis) with specially fine-tuned or tool-augmented subsystems. The advantage is that the general model doesn't need to be optimal in every domain, while vertical systems can achieve higher accuracy and efficiency within narrow domains. In actual deployment, the orchestration layer routes requests to appropriate models based on task type and enables collaboration through unified interface protocols (such as function calling). This layered strategy is also a key means of controlling inference costs — simple tasks are handled by lightweight models, and only complex tasks invoke the most powerful models.

Why "36 Hours" Is a Key Number

In the agent space, runtime duration is an underappreciated metric.

Most agent tasks in demos complete within minutes to hours. When runtime extends beyond 36 hours, the system faces an entirely different order of problems: state persistence, error recovery, resource leaks, goal consistency maintenance, and how to avoid cumulative drift over long time spans.

An agent system that can run stably for over 36 hours has effectively crossed the threshold from "demo" to "production." Such systems can take on real business processes — continuous monitoring, long-cycle data analysis, multi-phase project execution, and more.

Engineering Challenges of Large-Scale Concurrency

Running 10,000 agents simultaneously poses challenges beyond just the model inference layer.

Each agent needs to maintain its own state, context, and task progress. When the count reaches ten thousand, the orchestration system must solve problems including: coordination and conflict resolution between agents, dynamic allocation of computing resources, isolation and recovery of failed agents, and overall system observability.

Third Moment Research's ability to achieve this demonstrates that its underlying infrastructure is already quite mature. Capabilities at this scale have direct reference value for enterprise scenarios requiring large-scale AI agent deployment.

Agent Orchestration refers to the system layer that manages the lifecycle, task allocation, and coordinated scheduling of multiple agent instances. In ten-thousand-scale concurrent scenarios, the orchestration system typically needs several key capabilities: first, state management — each agent maintains independent runtime state (current task, historical context, intermediate results), requiring efficient distributed state storage; second, conflict resolution — when multiple agents operate on shared resources or produce contradictory outputs, an arbitration mechanism is needed; third, fault isolation — a single agent's crash or infinite loop must not affect other instances, typically achieved through sandboxed execution and heartbeat detection; fourth, observability — operations staff need real-time visibility into each agent's health status and output quality. The maturity of this layer often determines whether an agent system can scale from dozens to tens of thousands.

Implications for the Industry

Third Moment Research's practice points to a trend: the competitive focus for AI agents is shifting from "can it do it" to "can it keep doing it."

Single-task success rates are certainly important, but in real business scenarios, agents need reliability, persistence, and self-improvement capabilities. 36 hours of continuous operation and a concurrency scale of 10,000 fundamentally validate the feasibility of agents as "digital labor."

When agents can "go to work" like human employees — staying focused, producing continuously, and learning on the job — the imaginative possibilities of AI applications truly open up.