Inside OpenAI’s ‘self-operating’ infrastructure, where Codex-powered AI agents debug failures, manage releases, and compress the work of entire engineering teams (Photo by Samuel Boivin/NurPhoto via Getty Images)
NurPhoto via Getty Images
At OpenAI, a broken data pipeline no longer waits for an engineer to wake up; it triggers an AI agent. Instead of sitting in a triage queue, failures are investigated, debugged and, in some cases, resolved before a human even opens a dashboard. Releases move forward without manual orchestration, and on-call knowledge no longer lives in scattered Slack threads.
Inside OpenAI’s internal data platform — the system that powers everything from model training and safety pipelines to product analytics and financial reporting — software is beginning to operate itself at production scale. The platform supports more than 3,500 internal users and manages over 600 petabytes of data across roughly 70,000 datasets. Beneath that is a constantly moving architecture comprising of high-velocity Kafka streams, distributed Apache Spark jobs and orchestration layers coordinating thousands of workflows across systems.
Every prompt, every model iteration, every enterprise workflow touches this layer.
At the center are Codex-powered AI agents embedded directly into the infrastructure. They monitor pipelines in real time, tracking throughput, latency and data quality. When anomalies emerge — a stalled job, malformed event or schema drift—they trace and act. Sometimes that means restarting jobs or reallocating resources. In practice, it means generating fixes, validating them and preparing them for deployment.
“Our internal data agent(s) are not just getting a schema dump or a BI catalog export. It can draw on table definitions, ownership, documentation, query history, lineage, dashboards, permissions and the production code that generates the data,” Emma Tang, data platform lead at OpenAI, told me. “The big difference is that the model is operating on top of a data foundation we intentionally built. OpenAI has a unified data lake, canonical datasets, clean generating pipelines, code-defined table logic, maintained metadata, ownership, documentation, lineage and permissioning.”
The urgency comes from scale. Event volumes across OpenAI’s streaming systems have grown roughly 50 times in a year. At that level, dashboards break. Signals overwhelm human response loops. By embedding agents into the infrastructure itself, OpenAI is turning operations into a continuous process. Systems observe their own state, reason about what is happening and act in real time.
For data engineers, the shift is subtle yet profound. Their role moves from operating systems to supervising them. “The modern data analyst or data scientist should move up the stack,” Tang told me. “Agents should write the queries, find the right data sources, run the repetitive analysis, iterate on cuts and handle the manual work that slows people down.”
Codex, OpenAI’s cloud-based software engineering agent that began as a coding assistant, now behaves more like an execution layer. The agentic AI tool has grown to more than 3 million weekly users, with a significant share of activity extending beyond coding tasks into planning, documentation and operational work.
OpenAI engineers are building domain-specific agents on top of Codex that operate across streaming systems, data pipelines and machine learning infrastructure.
Tang explained that these data agents are different from coding agents. A coding agent has a relatively crisp context boundary: the repo. It can inspect files, run tests, see diffs and verify behavior. A data agent does not have that same boundary. Its “repo” is the company’s data foundation: the lake, pipelines, metadata, lineage, permissions, dashboards, metric definitions, owners and the operating knowledge around them. If that foundation is fragmented or poorly maintained, the agent has to reconstruct the company before it can answer the question.
She said this is where many earlier catalogs, BI tools and semantic layers fell short. They improved discoverability but often sat on top of fragmented data estates, with business logic hidden in notebooks or spreadsheets and multiple versions of the same metric. A better interface cannot collapse days into minutes if the underlying system does not know which source of truth to trust.
“Our internal data agent operates on top of a unified data platform. The lake, metadata, lineage, code, permissions, query execution, dashboards and notebooks are treated as one connected system,” Tang said. “So the agent can move through the same workflow a strong analyst would: identify the canonical table, inspect how it is produced, check ownership and documentation, reuse patterns from prior queries, run and repair SQL, interpret results and turn the answer into a durable artifact.”
She added that where it still breaks is also instructive. “It still inherits the limits of the data foundation, i.e., missing metadata, pipeline definitions missing from code, siloed data across systems.”
According to OpenAI, one internal release agent now manages updates to Apache Spark-based systems. It rolls out changes gradually, verifies stability over hours or days, generates pull requests and notifies teams for review. Likewise, another agent acts as an always-on on-call assistant. Instead of engineers digging through Slack threads, runbooks and past incidents, it retrieves relevant context — prior fixes, escalation paths and known failure modes — and applies it to new issues in real time. In development environments, agents spin up local services, launch browser sessions, test UI changes and validate behavior before a human reviews a single line. Engineers no longer spend hours verifying whether code works in practice.
Tang said a key part of a good response is making it easy to verify that the data is sound.
“Our system makes validation easier by exposing the artifacts needed for review. This includes assumptions the agent made, the chain of thought, the generated query, citations from internal references and confidence levels on answers,” she explained. “The agent also performs self-validation where possible, such as checking outputs against trusted ‘golden’ sources like verified dashboards and other independently reviewed sources of truth.”
Each agent builds on shared frameworks and composable components. Fixes become reusable, workflows are encoded and knowledge becomes persistent. This shows up in large-scale migrations. OpenAI has begun using Codex to generate hundreds of pull requests for these migrations automatically.
“Encoding operational knowledge into agents makes us more resilient if the knowledge stays explicit, versioned, tested and reviewable. It makes us fragile if Codex becomes an oral tradition with a better UI — where nobody knows why the workflow works or when it is stale,” said Venkat Venkataramani, VP of app infrastructure at OpenAI. “The goal should not be 'Codex remembers everything,’ rather ‘Codex can reliably find, execute and update the source of truth’.”
What makes that possible is context. Agents draw from multiple layers simultaneously, including metadata, historical queries, human annotations, code-level understanding of pipelines, institutional knowledge from internal tools and their own evolving memory.
“The trade-off is real,” said Venkataramani. “You reduce toil, but you raise the bar for the people designing the automation. I don’t think this means fewer people who magically do everything, but fewer people stuck doing repetitive operational work, and more engineers building durable systems and useful products.”
Recent updates to Codex reinforce that direction. The platform can operate computers directly — seeing screens, moving cursors and interacting with applications. It runs multiple agents in parallel, works across tools that lack APIs and continues tasks across days or weeks through persistent threads and memory.
Venkataramani said he does worry about opacity, noting that a system can be effective and still become dangerous if humans can no longer reason about it. “But this problem is not new. Every hyperscaler has thousands of engineers building thousands of systems with complex dependencies on codebases that are more than 10 years old. No single human in such environments can fully debug every novel failure mode.”
He said the answer is to make automation legible by documenting context for both humans and agents, including diffs, logs, decision traces, rollback points and post-mortems. “There is no substitute for humans with first-principles understanding of systems. The goal is to empower those humans with agents and context so they can mitigate and recover from truly novel failures faster.”
This shift toward agent-operated infrastructure is not happening in isolation. It is becoming the next battleground in the AI race.
Anthropic, OpenAI’s closest contemporary in frontier models, is advancing a parallel vision centered on agentic work. It's Claude Code and Cowork products position AI as a collaborator across the software lifecycle—writing code, reviewing pull requests, debugging issues, and iterating alongside developers. On benchmarks like
SWE-bench, Anthropic’s latest public models have pushed into the mid-to-upper 60% range on verified tasks, with internal variants reportedly higher.
However, OpenAI continues to challenge that narrative with its Codex-derived agents and broader agent-oriented benchmarks. On Terminal-Bench 2.0, OpenAI’s agent scores 77.3% versus Claude’s 65.4%. On general-purpose coding and agent-style tasks, Anthropic leads in SWE-bench and long-context reasoning, while OpenAI remains competitive on deterministic logic, factual retrieval, and multi-tool orchestration.
If OpenAI’s business model prevails, the execution layer becomes the control point. The company that owns the systems that actually run — deploying code, managing infrastructure, and resolving failures — defines how work gets done, often without direct user interaction. That distinction carries economic weight. Interface-layer dominance tends to drive usage and subscription revenue. Execution-layer dominance shapes operational dependency and long-term lock-in.
“People often think the hard part is connecting to Snowflake or Databricks and putting a model on top. That is the easy part. The hard part is making the company’s data reality legible to the agent: which tables are canonical, how metrics are defined, where business logic lives, who owns each dataset, what permissions apply, and more,” said Tang.
As infrastructure becomes more autonomous, a new constraint emerges: trust. OpenAI addresses this through continuous evaluation frameworks that compare agent outputs against expected results, using both SQL-level validation and data-level comparisons.
Tang said teams do not have to lose a deep understanding of their data as abstraction increases. “The parts being automated are mostly the mundane parts. That frees people to spend more time on higher-leverage work: asking better questions, understanding why the numbers changed, and deciding what to do next,” she said. “The risk is real only if teams treat agents as answer machines instead of reasoning partners.”
In AI’s next chapter, intelligence will still matter, but uptime may decide who leads the market. The advantage may shift to those who can help run these systems continuously, not just train them to peak performance.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。