Future AI Pulse
Quick answer: OpenAI’s Third Moment Research case study describes large numbers of agents working on investigations lasting more than a day. One useful lesson is simpler than the scale headline: repeatedly polling a slow operation can waste tokens and crowd a task’s context. Improve the underlying operation and preserve a clear job state before adding more agents.
What the case study reports
The October 14, 2026 account reports 10,000 concurrent agents and research sessions lasting 36 to 48 hours. In one example, agents waited on a backtest that took approximately three hours. OpenAI says Codex identified caching and memory-packing changes that reduced it to approximately three to five minutes.
The account also describes an outer harness that decides whether unfinished work should be retried, terminated or escalated. These are vendor-published customer claims about a particular environment, not an independent benchmark or a guarantee that an ordinary chat will run continuously. Future AI Pulse has not reproduced the system.
Waiting is a state worth representing
A long task needs a way to distinguish useful computation from checking progress. If an agent receives the same “still running” response repeatedly, those messages can consume attention without adding evidence. A job record should tell the system what is running, what would count as completion and where the result will appear.
For a small workflow, keep a stable job identifier, the input version, the submission time and the expected artifact. The agent can refer to that record when it resumes instead of rebuilding the entire history from repeated status messages.
Separate progress checks from result checks
A progress signal establishes that something is happening. A completion signal establishes that the worker stopped. Neither automatically establishes that the output is correct. Inspect the artifact against the original acceptance criteria after completion.
For example, a batch report can finish successfully while omitting one input file. Check expected file count and required sections before marking the task complete. Likewise, a benchmark run can return an exit code without producing the measurements a reviewer needs.
A bounded job-management brief
Track this job using its stable identifier and input version. Record the expected output and acceptance criteria. Use available completion notifications before repeated status polling. If polling is required, use the configured interval and retry limit. When the job finishes, inspect the actual artifact. Do not resubmit an ambiguous job until its current state is checked. Escalate missing output or persistent failure with the specific evidence.
This is an original workflow brief, not a tested command for Third Moment’s system. The interval and retry count should be configured in the real scheduler or worker. A prompt alone does not supply persistent infrastructure.
Investigate the slow operation first
Use timing data to identify whether the job spends time loading inputs, recomputing unchanged work, waiting for a resource or producing its result. Change one part at a time and compare representative inputs. Check that the optimization preserves the output you care about.
Caching needs an invalidation rule. If an input or configuration changes, a previous result may no longer be valid. Record the key used to identify reusable work and include a case that should force recomputation in your verification.
Keep restart decisions explicit
A restarted task should know whether it is continuing a prior job or starting a new one. Preserve useful intermediate artifacts, but do not silently reuse uncertain results. Retrying forever can multiply cost without resolving the actual blocker.
Our GPU scheduling coverage examines queue delays, while the voice-agent task brief discusses clear stopping conditions. Together, these ideas help make long work easier to inspect.
Illustrative server photograph: Kevin Ache / Unsplash, under the Unsplash License. Not Third Moment infrastructure.
Sources
Source checked October 14, 2026. Written with AI assistance; operational examples are editorial suggestions, not reported testing.