---
type: "post"
id: "how-third-moment-research-runs-10-000-agents-f1f301"
slug: "how-third-moment-research-runs-10-000-agents-f1f301"
title: "How Third Moment Research runs 10,000 agents for 36+ hours"
description: "OpenAI’s account of Third Moment Research describes a system where the harness, not a single uninterrupted model run, carries long-running research work."
retrieval_nugget: "OpenAI’s account of Third Moment Research describes a system where the harness, not a single uninterrupted model run, carries long-running research work."
published_at: "2026-10-14"
updated_at: "2026-10-14"
record_date: "2026-10-14"
date_kind: "published_at"
topics: []
source_urls: ["https://openai.com/index/third-moment-research"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:how-third-moment-research-runs-10-000-agents-f1f301"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:0f87352b2125155aa6ed892378332acf529fb6ee5753e437b3c3a88ca3fe8734","reviewed_at":"2026-10-14T16:11:11.618Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Thesis: Durable agent work depends on an outer runtime that can inspect progress, recover from incomplete attempts, and route work onward.","observed_facts":[{"text":"OpenAI says Third Moment Research runs 10,000 agents concurrently and autonomous research sessions last 36–48 hours.","source_urls":["https://openai.com/index/third-moment-research"]},{"text":"The account says an outer harness reviews unfinished Codex trajectories and can retry, terminate, or escalate work for human review.","source_urls":["https://openai.com/index/third-moment-research"]},{"text":"OpenAI reports a backtest optimization from approximately three hours to approximately 3–5 minutes after caching and memory-packing changes.","source_urls":["https://openai.com/index/third-moment-research"]}],"mechanism":"Mechanism: Codex operates as a subagent inside an outer harness, which examines an unfinished trajectory and can retry, terminate, or escalate it for human review.","why_now":"Why now: The case ties current long-horizon model capability to the runtime controls needed to keep expensive parallel work useful rather than merely continuous.","implications":["First implication: Production agent architectures should treat orchestration and recovery as core system components.","At large scale, environment security validation belongs in the operating loop."],"evidence_boundary":"Evidence boundary: OpenAI reports these methods and results, but the supplied source does not independently verify them or establish that they generalize beyond this setup.","watch_conditions":["First watch condition: Whether future deployments publish comparable measures of successful task completion, escalation frequency, and containment failures, alongside agent counts and runtime."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this case starts a distinct analytical record about long-running research agents and their outer runtime controls."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/how-third-moment-research-runs-10-000-agents-f1f301/","markdown":"https://newruntime.com/posts/how-third-moment-research-runs-10-000-agents-f1f301.md","json":"https://newruntime.com/posts/how-third-moment-research-runs-10-000-agents-f1f301.json"}
---

# How Third Moment Research runs 10,000 agents for 36+ hours

## Retrieval answer

OpenAI’s account of Third Moment Research describes a system where the harness, not a single uninterrupted model run, carries long-running research work.

OpenAI’s account of Third Moment Research describes a system where the harness, not a single uninterrupted model run, carries long-running research work. Thesis: Durable agent work depends on an outer runtime that can inspect progress, recover from incomplete attempts, and route work onward. This is a company-published case study, so the figures should be read as reported evidence rather than independently audited outcomes.

Mechanism: Codex operates as a subagent inside an outer harness, which examines an unfinished trajectory and can retry, terminate, or escalate it for human review. OpenAI says Third Moment Research runs 10,000 agents concurrently and autonomous research sessions last 36–48 hours. In one backtesting example, caching and memory packing cut a run from approximately three hours to approximately 3–5 minutes. Why now: The case ties current long-horizon model capability to the runtime controls needed to keep expensive parallel work useful rather than merely continuous.

First implication: Production agent architectures should treat orchestration and recovery as core system components. The source also reports Daybreak testing whether agents can break out of their simulation environment and reach the wider network, making security validation part of operating at scale. Evidence boundary: OpenAI reports these methods and results, but the supplied source does not independently verify them or establish that they generalize beyond this setup. First watch condition: Whether future deployments publish comparable measures of successful task completion, escalation frequency, and containment failures, alongside agent counts and runtime.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)