Third Moment Research Agents: 10,000 at Once

Third Moment Research agents run 10,000 at a time in 36–48-hour sessions, according to OpenAI’s October 14 case.

Third Moment Research agents: completion needs evidence

The official case says its Sol/Sol 6.1 outer harness checks incomplete Codex work and can retry, terminate, or escalate to human review.

For a smaller implementation, our proposed starting point is an explicit completion condition. Consider an original, untested research task: compare two data-processing methods and save both timing results alongside a check that their outputs match. If the agent writes a confident summary after measuring the first method, the summary is not the missing second result. A supervisor needs to inspect the requested artifacts before accepting the task as complete.

The same missing result can have different causes. The second process might still be running, the agent might have forgotten it, or the input might have changed since the first measurement. Each situation calls for a different next action. Simply sending “continue” treats those states as interchangeable and can launch duplicate work or compare results that no longer share an input.

Keep the final message and the completion record separate. In this example, the record would name both result files, the input revision, and the output comparison. The message explains the conclusion; the record makes it possible to check whether the conclusion has the evidence the task required.

Waiting, resuming, and retrying are different operations

Our illustrative supervisor would first check whether the missing measurement has a live process or an already saved result. A running calculation should stay in a waiting state. An existing result should be inspected before another copy is started. An absent result can justify resuming the unfinished step, provided the input and task conditions still match.

A retry is appropriate for a failed execution, not automatically for an unreported success. Suppose a worker saved its timing result but lost its connection before delivering the final message. Running that measurement again consumes resources and may produce a different timing. Recovering the saved result answers a different question: what did the original execution actually finish?

For a changed input, neither waiting nor repeating the last instruction repairs the comparison. In our example, the owner would decide whether to rerun both methods against the new revision or keep the earlier experiment intact. That decision belongs in the supervisor’s visible record so later readers know which comparison they are examining.

This is an application-level interpretation of the case, not Third Moment’s disclosed internal schema. It is useful for repository owners and research-tool builders because it makes an early stop diagnosable. The owner can see whether work remains, whether a dependency is still running, and whether a human decision is needed before consuming another research cycle.

Separate tool runtime from status-check traffic

An earlier GPT-5.6 optimization cut a backtest from around three hours to three to five minutes. Repeated completion checks also filled context while tools worked.

An illustrative count shows why that distinction matters. A three-hour process checked once per minute produces 180 checks. Changing the interval to five minutes produces 36 checks over the same duration. Those invented intervals do not reconstruct Third Moment’s operation, and fewer checks alone do not make the underlying calculation faster. They change how much status traffic surrounds it.

For the two-method example, record a timeline with the process start, any result-file creation, and the agent’s next substantive action. A supervisor can then tell whether a delay came from computation, from waiting after a result was available, or from repeating a completed measurement. That is more actionable than one total duration labeled “agent speed.”

OpenAI’s latency optimization guide distinguishes inference speed from request overhead and application work. It also explains that independent steps can run in parallel, while dependencies constrain execution order. These distinctions support measuring the waiting path before choosing a faster model.

For a separate case focused on retained browser history and cached input, see the Asana workflow report. Here, the question is whether a supervisor is asking for new reasoning when it only needs to learn that a tool has finished.

Preserve the evidence needed after context compaction

A task can continue through compacted context without giving a human reviewer a readable completion ledger. OpenAI’s compaction documentation describes an opaque item that carries key prior state and reasoning into later turns using fewer tokens. It is not intended to be human-interpretable.

For our proposed supervisor, maintain readable task evidence separately: the requested comparison, the input revision, the saved result locations, and the remaining dependency. That record is not a replacement for the model’s context mechanism. It serves the different job of letting an operator inspect the state without interpreting an encrypted compaction item.

One checkpoint in the example might say that the first measurement is saved, the second is running, and the output-equivalence check cannot begin until both finish. After a context change or an interrupted connection, the next action is constrained by that checkpoint. Starting two fresh measurements would be visibly inconsistent with the recorded state.

When using the Responses API, the official compaction guide gives distinct continuation patterns for input-array chaining and previous-response chaining. It specifically says not to manually prune when using the latter. Choose the documented pattern for the implementation instead of treating every shortened history as equivalent. This API guidance explains the mechanism; it does not establish which configuration Third Moment uses.

The GPT-6 Astra profile and GPT-6.1 Sol page provide separate model references. A durable completion record remains an application responsibility whichever model a team evaluates for the work.

Check the environment before extending autonomy

Daybreak tests environment isolation before simulations, OpenAI reports. Larger collaborating fleets remain a plan, not a reported deployment.

For the proposed comparison task, define where the worker may write results and which resources it may access. An incomplete job should not silently gain broader permissions when the supervisor retries it. A missing input that requires new access is a reason to request a decision, not evidence that the agent should work around the boundary.

This connects the early-stop problem to supervision: the next action must follow both the saved execution state and the permissions already granted. The transferable lesson is not to make an agent run indefinitely. It is to make waiting, recovery, completion, and escalation distinguishable from the evidence the task actually produced.