In this briefing
  1. 01OpenAI Turns Performance Alerts into Tasks with External State
  2. 02Modal Places Improvement Experiments in Monitored Isolated Environments
  3. 03Microsoft Turns Production Records into Evaluation-Gated Candidate Improvements
  4. 04All Three Implementations Keep the Final Decision Outside the Model
  5. →What to watch next
  6. ↗Sources and verification
Key points
  1. OpenAI places performance alerts, root-cause investigation, independent adversarial review, proposed fixes, engineering review and post-deployment recovery verification in an explicit state workflow. The company says that the workflow tracked more than 1,700 investigations in just over three months and found the root causes of more than 600 cases.
  2. Modal's published customer case places the launch, attempt, verification and reporting of reinforcement-learning tasks in isolated environments. Abundant runs more than 2 million isolated environments per month as of month 11 of 2026, but the scale and reliability figures come from public statements by the vendor and customer.
  3. Microsoft Foundry connects production traces, expert feedback, candidate optimisation, pre-release statistical testing and continuous evaluation of production traffic. Microsoft also states that enterprise teams decide which changes reach production, while several new capabilities remain in preview.
Signal 01

OpenAI Turns Performance Alerts into Tasks with External State

OpenAI announced Performance Factory on 19 November. The system comprises a task-tracking service, a dispatch loop, a cloud Codex runtime interface and event subscriptions. Engineers use configuration files to define task states, permitted state transitions and the steps that require an engineer, while Agents investigate metrics, logs, traces, processor profiles, memory snapshots, deployment changes and relevant code after a performance alert fires. The task stores its state and key findings outside any one Agent's context, so other Agents or engineers can continue the same problem after code review or deployment.

OpenAI says that the first performance-regression workflow tracked more than 1,700 investigations in just over three months and found the root causes of more than 600 cases, averaging 17 automated investigations per day. OpenAI also explains that an independent Codex session tries to disprove the initial root-cause conclusion, while engineers review proposed fixes and decide whether to merge the code. After a fix is deployed, another Agent checks whether performance has returned to the original baseline; if performance has not recovered, the workflow returns to the investigation stage. The figures come from OpenAI's internal system records, and OpenAI has not provided an external independent audit in the public material.

What this may mean for enterprise adoption

For enterprise implementation, an enterprise needs to store Agent runtime state and evidence outside the model conversation, and the enterprise needs to define explicit state transitions for alerts, investigation, review, code changes, human approval, deployment and recovery verification. Independent review can reduce the risk that an Agent stops at the first plausible explanation, but independent review cannot replace the judgement of an engineer who understands the business and system architecture. The enterprise also needs to restrict Agent access to production-observability data and code repositories, and the enterprise needs to record every action performed by the service identity.

Signal 02

Modal Places Improvement Experiments in Monitored Isolated Environments

Modal published Abundant's reinforcement-learning runtime architecture on 19 November. Abundant designs tasks that contain instructions, a goal and a verifier, and each task runs in a Modal Sandbox. Modal says that Abundant has created more than 75,000 rollouts and that Abundant runs more than 2 million isolated environments per month as of month 11 of 2026. Modal Functions start tasks, clean up unresponsive runs and review results, while the isolated environments host the model's actual attempts.

Abundant divides each rollout into launching the task, attempting the task, verifying the work and reporting the result. A monitor inside the isolated environment sends a heartbeat every 30 seconds, as well as live events and the final result; if the isolated environment stops responding, the control plane tries to connect 3 times before terminating the environment. Modal also lets an isolated environment temporarily increase its central-processing-unit and memory use when the underlying machine still has spare resources. The mechanism can reduce idle resources reserved for peak demand, but Modal's public material does not provide an independent comparison with other platforms under the same task, failure and pricing conditions.

What this may mean for enterprise adoption

For enterprise implementation, an enterprise needs to fix the task inputs, runtime image, resource boundaries, verifier, event records and termination conditions for each improvement experiment. An isolated environment can reduce state contamination between experiments, while heartbeat and unresponsive-run handling can help the control plane distinguish a normal long-running task from a failed run. Elastic resources can satisfy peak demand only when the underlying machine still has spare capacity, so an enterprise cannot treat elasticity as a guaranteed capacity commitment. The enterprise still needs to check concurrency limits, failure recovery, data retention, network permissions and cost per task.

Signal 03

Microsoft Turns Production Records into Evaluation-Gated Candidate Improvements

Microsoft announced the Microsoft Foundry continuous-improvement process on 17 November. The process establishes an enterprise's own quality standard, identifies problems in runtime records, generates candidate optimisations and then validates whether the optimisation passes the existing standard. Microsoft Foundry records the task that an Agent received, the steps that the Agent performed and the answer that the Agent produced, while experts can annotate real conversations and the platform can convert those annotations into evaluation datasets and a fixed test set. Trace annotations, evaluation datasets and the fixed test set remain in preview.

Microsoft says that Agent optimizer changes Agent instructions, skills or model choice, and the tool tests candidate changes against enterprise evaluation standards before the changes enter production. Problem insights are generally available, while one-step candidate fixes from production problems, skill optimisation and cost optimisation remain in preview. Microsoft Foundry can already sample and score production traffic continuously, and the platform can rerun fixed test sets and safety tests on a schedule. Microsoft explicitly states that enterprise teams decide which changes enter production, so a candidate generated by the platform does not mean that a production version has been approved.

What this may mean for enterprise adoption

For enterprise implementation, an enterprise first needs to define evaluation standards that reflect business quality, safety, latency and cost before the enterprise can decide whether a candidate change genuinely improves the system. Production records can help an enterprise find real problems, but production records can also contain sensitive inputs, tool arguments and business results, so the enterprise needs to restrict the recording scope, access permissions and retention periods. Preview capabilities can support controlled experiments, but preview capabilities cannot count in advance as a stable production contract. The enterprise also needs to verify statistical significance, business impact and rollback capability separately.

Signal 04

All Three Implementations Keep the Final Decision Outside the Model

OpenAI uses an independent session to challenge a root-cause conclusion, while engineers decide whether to merge a fix; Abundant uses a verifier and a reviewer Agent to check task results and evaluation awareness; Microsoft Foundry tests candidate changes against an enterprise-defined standard, while the enterprise team decides which changes enter production. The three implementations serve different scenarios, but the same model that proposed an improvement does not prove by itself that the improvement has succeeded.

The three sets of material also expose the operating limits of an improvement loop. OpenAI states that proposing a fix is harder than finding a root cause and that some proposals may be too complicated or introduce new problems. Modal's elastic resources depend on spare capacity in the underlying machine, and verifiers need to stop models from obtaining passing results through shortcuts. Several Microsoft Foundry capabilities for annotation, optimisation and fixed test sets remain in preview. The public material from OpenAI, Modal and Microsoft Foundry does not provide a comparison of quality, latency, privacy, cost and failure recovery on the same enterprise task.

What this may mean for enterprise adoption

For enterprise implementation, a verifiable improvement loop needs at least independent state, a reproducible environment, explicit evaluation, review by different roles, human release approval and post-deployment result verification. An enterprise cannot treat a model-generated root-cause explanation, code fix or evaluation summary as a completed business result. The enterprise needs to retain the links between the original production signal, candidate change, evaluation version, approver, deployment record and post-deployment metrics, and the enterprise needs to return the workflow to investigation when the evidence is insufficient.

Verification

Sources and verification

  1. How OpenAI’s Performance Factory keeps ChatGPT fastOpenAI · 2026-11-19 · Official documentation
  2. Continuous improvement comes to Microsoft FoundryMicrosoft · 2026-11-17 · Official announcement

Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.

← Back to AI Daily Briefing