Third Moment Research Agents Run for 36 Hours, but Scale Raises a Harder Question

Third Moment Research agents now run 10,000 at once, with autonomous investigations lasting 36 to 48 hours. That scale moves the company far beyond the short, closely supervised agent sessions common in business software.

The startup says GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, Codex, and Daybreak support this system. Together, they let agents plan experiments, use tools, create subagents, preserve context, and respond when work stalls.

Yet the important contest is not Third Moment against another AI laboratory. It is sustained, measurable research against the failure patterns that make long-running agents unreliable. Those include premature completion, context loss, wasted computation, security exposure, and experiments that produce noise instead of useful findings.

OpenAI published the customer profile on October 14, 2026. Nearly every performance claim comes from Third Moment or OpenAI, so the results should be read as reported operational evidence, not an independent audit.

The case still offers a detailed view of where autonomous agents are heading. The emerging unit of AI work is no longer one response or one coding task. It is a managed investigation that consumes data, software, model calls, and many dependent experiments over several days.

Third Moment Research Agents Turn One Task Into Thousands of Experiments

The headline number matters because 10,000 concurrent agents represent a different operating model, not merely a larger chatbot deployment.

Third Moment builds systems that find recurring agent failures and generate synthetic scenarios around them. Synthetic scenarios are constructed test cases designed to reproduce a weakness without waiting for the same failure in live traffic.

The company then uses those scenarios to improve its agents. New problems feed back into the process, creating a recurring cycle of observation, testing, and adjustment.

Third Moment’s own site describes a related endpoint model. Customers send real workflow traffic through an OpenAI-compatible endpoint. The company groups requests into trajectories, identifies failure patterns, builds evaluations, and modifies the model or harness after customer approval.

A trajectory records the agent’s actions and the environment’s responses across a task. That structure matters because a failed outcome rarely reveals which earlier decision caused the problem.

The OpenAI profile says Third Moment runs 10,000 agents concurrently. Individual research sessions continue for 36 to 48 hours, and one day of operation can consume one trillion OpenAI tokens.

Those figures describe three separate scaling problems. Concurrency tests infrastructure, long sessions test continuity, and token volume tests economic efficiency. Improving one does not automatically solve the others.

Third Moment’s initial application focuses on quantitative finance. The domain offers historical data, repeatable simulations, and measurable results, making it useful for studying agent behavior.

Agents receive a terabyte of data, documentation, and access to a research repository. The repository includes modules for feature engineering, model training, trading strategies, and market simulation.

Their job is not simply to generate code. They modify a baseline trading system, test strategies on historical data, evaluate whether results are meaningful, and choose the next experiment.

That closed loop gives the agents feedback after every attempt. A weak result can redirect the investigation, while a promising result can trigger deeper tests or parallel work.

The setup also provides a clearer score than many knowledge-work tasks. A trading simulation can measure runtime, correctness, stability, and historical behavior. A general research memo often lacks such direct feedback.

Third Moment says its agents have improved the baseline system. However, OpenAI’s profile does not publish investment returns, risk-adjusted performance, or out-of-sample validation.

That omission does not erase the engineering result. It does limit what readers can conclude about the financial value of the research.

The central change is therefore operational. Third Moment has assembled an environment where agents can continue a structured investigation across days, rather than ending after one model response.

That shift creates the article’s main tension. Scaling agent activity is easy to count, but sustained useful progress is much harder to establish.

Why Long-Running Agents Put the Harness Under Pressure

A model can be highly capable and still fail a long investigation if the surrounding harness loses goals, context, or control.

An agent harness is the software layer that manages model calls, tools, state, approvals, and task progress. Third Moment uses the Codex harness inside Dockerized environments for its quantitative research.

GPT-6 models provide reasoning, while Codex supplies the operating loop. Agents can create plans, set objectives, call tools, spawn subagents, and compact context as the session grows.

Context compaction reduces accumulated conversation history while trying to preserve the information needed for later work. It prevents a long session from exceeding the model’s active context window.

OpenAI’s explanation of the Codex agent loop shows why this layer matters. Codex tracks conversation state and replaces oversized input with a smaller representation when necessary.

Compaction is useful, but it is lossy by design. A compressed record can omit an earlier constraint, obscure why a decision was made, or overemphasize recent activity.

Long-running agents also face a second problem: they can decide that an unfinished task is complete. Third Moment responds by placing Codex inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol.

When an agent stops early, the outer system examines its trajectory. It then decides whether to retry the job, terminate it, or send it for human review.

This arrangement separates task execution from supervision. The worker agent pursues the research, while the outer harness judges whether its behavior still serves the original objective.

That distinction becomes more important as sessions grow. An agent can produce many valid intermediate steps while drifting away from the question it was meant to answer.

The same problem appears elsewhere in the industry. Anthropic’s long-running agent guidance says compaction alone does not guarantee consistent progress across multiple context windows.

Anthropic describes agents trying to do too much at once and struggling to preserve continuity between sessions. Its approach emphasizes incremental work and persistent artifacts that help the next session recover state.

Third Moment’s approach differs in scale and application, but the underlying lesson is similar. Long-duration autonomy depends on external structure, not only a larger model.

The Codex harness also provides subagents, which let a parent process divide an investigation into narrower tasks. This can increase throughput when experiments are independent.

However, more subagents also create more coordination work. Their results must be compared, rejected, combined, and connected to the original objective.

The parent must know which experiment changed the evidence and which merely repeated an earlier attempt. Otherwise, parallelism produces activity without advancing the investigation.

OpenAI has reported that harness design can materially affect model performance. Its Codex platform describes retained reasoning and context compaction as core parts of a reusable agent loop.

That makes Third Moment’s 36-to-48-hour sessions a test of orchestration as much as intelligence. Model quality helps, but the harness decides whether that quality survives thousands of actions.

How Third Moment Agents Cut Waiting Without Losing the Research Thread

The clearest reported gain came from removing wasted waiting, which gave agents more time and context for actual experiments.

Third Moment recorded thousands of simulation trajectories, including agent actions, environment responses, timestamps, and relevant code. It then used Codex with GPT-5.6 to inspect that evidence.

The analysis identified a bottleneck in backtesting. Testing one strategy across one year of order-level US equity data took about three hours.

During that delay, agents repeatedly checked whether the job had finished. Those status calls consumed tokens and filled the active context with information that added little research value.

This behavior illustrates a common automation trap. A slow external process does not merely delay the answer. It also changes how an agent spends its limited attention.

Codex reportedly profiled the backtest, examined its inputs, and identified two changes. One added caching, while the other packed data into memory more efficiently.

Third Moment says the resulting system reduced the backtest from roughly three hours to about three to five minutes. That is approximately a 97 percent reduction in runtime.

The company also says Codex tested the changes and deployed the fix into production. OpenAI does not provide the implementation, benchmark protocol, or an independent reproduction.

The result remains useful because it connects infrastructure performance directly to agent behavior. Faster simulation means fewer status checks, less context pollution, and more experiments within the same session.

It also shows how Third Moment expects agents to improve the environment they use. Codex was not only a researcher inside the system. It became a tool for changing the system’s own constraints.

This self-improvement claim needs careful definition. The agents are not independently rewriting every part of their operation without oversight.

The documented process uses recorded failures, engineered test environments, measurable evaluations, and production controls. The company also says customers approve which identified issues should be fixed.

That is closer to an evaluation-driven optimization loop than unconstrained recursive self-improvement. The distinction matters because the second phrase suggests much broader autonomy than the evidence supports.

Model selection is another part of the mechanism. Third Moment evaluates models across optimization, machine learning systems, and quantitative research.

The company tests frontier models at high reasoning settings. It measures their ability to navigate repositories, write efficient code, and continue investigations.

According to co-founder Jeffrey Yan, GPT-6.1 Sol became the company’s most-used model because of its token efficiency. He said it delivered performance similar to GPT-6 Astra while using fewer tokens.

OpenAI describes GPT-6.1 Sol as approaching Astra-level performance on coding, computer use, and professional tasks. That comparison comes from OpenAI’s evaluations, not Third Moment’s independent benchmark.

Third Moment reports one internal benchmark that asks agents to reduce the update latency of a live scoreboard service. A solution must preserve outputs and run on the same hardware.

The team calculates speedup by comparing a model’s solution with an expert-crafted solution. Co-founder Adam Khakhar says GPT-6 Astra delivered the largest speedup on that task.

For narrower work, the company uses more efficient models such as GPT-5.6 Terra. One example is checking the simulation environment for specific security weaknesses.

This routing strategy matters more than any single model ranking. A 36-hour investigation can include planning, coding, simulation, security testing, and repetitive checks.

Using the most capable model for every action would waste resources. Using a smaller model for every action could weaken the decisions that shape later experiments.

Third Moment instead assigns different models according to the work. That makes the system a portfolio of reasoning capabilities managed by a common harness.

The Real Contest Is Sustained Research Against Agent Drift

Third Moment’s strongest claim is not that more agents are better, but that its agents can preserve a coherent investigation while evidence keeps changing.

Khakhar says GPT-6.1 Sol can maintain one computational research thread across several days without losing the plot. This is a company statement, and no independent continuity evaluation accompanies it.

The OpenAI profile provides examples of longer reasoning. With GPT-5.5 and GPT-5.6, agents reportedly began using experimental results to design follow-up tests.

With GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol, Third Moment says agents can investigate root causes through hundreds of parallel experiments. Subagents explore individual branches while the broader system compares their findings.

In one reported session, GPT-6.1 Sol examined non-displayed trades on an exchange. Non-displayed trades execute without showing the order in the public order book beforehand.

The agent identified that these trades represented an unexpectedly large portion of volume. Khakhar said the human research team had not initially considered that result.

This example supports the idea that an agent can uncover an overlooked research direction. It does not establish whether that observation produced a durable trading signal.

The difference between discovery and validation is crucial in quantitative research. Historical datasets contain many correlations that disappear when conditions change or transaction costs enter the calculation.

Running hundreds of experiments expands the search space. It also expands the number of chances to find an apparently strong result by accident.

A credible research system therefore needs controls for overfitting, leakage, multiple testing, and changing market conditions. OpenAI’s profile does not describe Third Moment’s procedures for those risks.

The company says agents assess whether improvements reflect meaningful findings or noise. Yet it does not publish the thresholds, holdout design, or review process behind that judgment.

This verification gap is the most important skeptical angle. The scale figures are concrete, but the article offers limited evidence about research quality after those agents finish.

Ten thousand agents can repeat similar assumptions at high speed. They can also amplify a flawed objective or exploit a weakness in an evaluation.

The outer harness can catch jobs that stop early, but premature stopping is only one failure mode. An investigation can continue for days while pursuing the wrong proxy.

It can also preserve internal coherence while drifting from reality. A chain of logically connected experiments is not useful if the original data or evaluation is misleading.

Human review therefore remains part of the architecture. Third Moment’s public description says customers review identified issues before the company fixes them.

The outer harness can also escalate unfinished work. These controls suggest that the system manages autonomy through boundaries instead of removing people entirely.

That is consistent with a broader pattern in long-running agents. Humans increasingly define objectives, permissions, evaluations, and escalation rules while agents perform more intermediate work.

The pressure falls on teams that still treat agents as isolated chat sessions. They will need environments where tools, documentation, permissions, and results remain legible throughout a long task.

For engineering groups, that includes maintaining an engineering knowledge base. An agent cannot reliably use decisions or constraints that remain scattered across inaccessible systems.

The emerging competition is therefore between operating systems for agent work. Model benchmarks still matter, but environments, evaluations, and context management increasingly determine whether performance survives contact with real tasks.

Daybreak Becomes a Security Gate for 10,000 Autonomous Workers

At this scale, security testing becomes part of the agent runtime because one escaped process can turn experimental access into operational exposure.

Third Moment runs its quantitative agents inside Dockerized environments. Containers can isolate software and files, but isolation depends on configuration, permissions, and the surrounding infrastructure.

The company uses OpenAI Daybreak to test those environments before simulations begin. Its agents perform penetration testing against the simulation setup and supporting infrastructure.

The tests ask whether an agent can escape its assigned environment or reach the wider network. Third Moment says it could not test this exposure autonomously at scale without Daybreak.

OpenAI describes Daybreak controls as a governed system for vulnerability validation, risk prioritization, remediation, and security evidence. Access is designed for approved defensive work.

The Third Moment use case is notable because the object being tested is another agent environment. AI systems are helping evaluate whether AI workers can cross their assigned boundaries.

That creates a layered design. One set of agents conducts research, a security system tests its workspace, and an outer harness monitors whether work should continue.

Each layer addresses a different risk. Containers constrain execution, Daybreak probes the boundary, and the outer harness handles task-level failure or escalation.

None of those layers guarantees safety. Penetration tests cover known scenarios and discovered weaknesses, while new configurations can introduce new paths.

Autonomous scale also increases the importance of credentials and network policy. A small mistake repeated across 10,000 workers can create a much larger exposure than one supervised session.

The OpenAI profile does not specify whether every agent receives identical permissions. It also does not describe network architecture, credential lifetime, audit retention, or incident response.

Those omissions are reasonable for a public customer story. They prevent readers from judging whether Third Moment’s security architecture matches the size of its deployment.

There is also a tension between research freedom and containment. Agents need tools, data, and computational access to investigate problems, but every added capability enlarges the attack surface.

Tight restrictions can make an agent ineffective. Broad permissions can turn an erroneous or manipulated action into a serious incident.

Prompt injection remains another concern. An agent processing external or generated content can encounter instructions that conflict with its assigned task.

Synthetic scenarios can test such behavior, but successful tests do not cover every future input. Long sessions provide more opportunities for malicious or confusing content to enter the trajectory.

Compaction can also affect security. If an earlier safety constraint disappears from the active context, the agent may behave differently later in the session.

Infrastructure-level enforcement therefore matters more than textual instructions alone. Network rules, sandbox policies, scoped credentials, and approval gates should continue working even when the model’s context changes.

Third Moment’s Daybreak use suggests the company recognizes this requirement. Still, the public evidence describes a testing process, not a complete safety evaluation.

The relevant benchmark is not whether an agent escaped during one campaign. It is whether the system continuously detects new paths as models, tools, data, and infrastructure change.

What Will Show Whether the 10,000-Agent Model Actually Works

The next test is whether Third Moment can convert extraordinary activity into reproducible gains without weakening oversight or research quality.

The first signal is independent validation of research outcomes. Published evaluations should separate infrastructure improvements from financial or scientific discoveries.

The backtest optimization offers a concrete engineering result because it compares runtime before and after a defined change. Comparable evidence for research quality would strengthen the larger story.

Useful reporting could include blinded evaluations, held-out tasks, reproducibility rates, and the share of agent findings that survive human review. These measures would show whether longer sessions add value.

If Third Moment publishes such results, the case for sustained autonomous research becomes stronger. If evidence remains limited to internal anecdotes, the 10,000-agent figure will remain difficult to interpret.

The second signal is expansion beyond quantitative finance. Third Moment identifies supply chains, large software repositories, and document-heavy business processes as potential applications.

These domains offer less immediate feedback than a backtest. A supplier recommendation or contract discrepancy may require judgment that cannot be reduced to one numerical score.

Successful deployment there would show that the system can work with ambiguous evidence and organizational constraints. Failure would suggest that its current approach depends heavily on simulation-friendly environments.

The company’s public product direction already points toward customer-specific endpoints. It aims to diagnose real workflow failures and improve models or harnesses against approved evaluations.

That approach can transfer across industries if customers provide representative trajectories and reliable acceptance criteria. It becomes weaker when success remains subjective or changes across teams.

The third signal is progress toward Third Moment’s stated roadmap. The company plans to move from individual investigations to hundreds of thousands of collaborating subagents.

It also envisions trillions of tokens supporting chains of dozens of dependent experiments. At that point, coordination quality will matter more than raw parallel capacity.

A dependent experiment cannot start meaningfully until an earlier result is trustworthy. Errors near the beginning of the chain can influence every later branch.

Watch for evidence that the orchestration system rejects weak findings, revisits assumptions, and records why one path was chosen over another. Those capabilities would strengthen the company’s central claim.

Also watch the rate of human escalation. A falling escalation rate can signal improvement, but only if task quality and safety remain stable.

If escalations fall because the system becomes less cautious, the same metric would hide risk. Third Moment will need paired measures rather than one headline number.

The model mix will matter too. GPT-6.1 Sol currently occupies a central role because Third Moment reports a favorable efficiency-to-performance balance.

Future model releases can change that calculation. A better system should allow models to be replaced without rebuilding every workflow or invalidating historical evaluations.

That favors a harness-centered architecture. The model supplies reasoning, while the surrounding system preserves goals, tools, evidence, permissions, and recovery behavior.

For developers and enterprise buyers, the practical lesson is not to copy the 10,000-agent count. Most organizations do not yet have tasks, evaluation systems, or infrastructure that justify such scale.

Start by asking whether a workflow has observable outcomes, recoverable state, controlled tools, and clear escalation rules. Without those elements, adding agents multiplies uncertainty.

Knowledge workers should care for a related reason. Long-running agents can search more material and test more possibilities, but their conclusions still depend on accessible, current, well-structured context.

Third Moment Research agents offer a preview of AI work measured in days instead of prompts. The decisive question is whether that endurance produces findings people can verify and trust.

Over the next three months, watch for independent evaluations, deployments outside quantitative finance, and evidence from larger collaborating-agent trials. Those signals will reveal whether this architecture scales useful judgment or only computational activity.

For teams exploring autonomous research, the next action is simple: choose one measurable workflow and document its failure conditions before adding more agents. If a system cannot explain what success means, when to stop, and when to escalate, longer runtime will not fix it. It will only make the uncertainty more expensive.