On 2 November 2026, OpenAI published an engineering post that would have been unthinkable two years ago. The company had built a new backend for Git repositories called Pando because its existing source control system was collapsing under the sheer volume of work generated by Codex, its AI coding agent.

The numbers are striking. Commits to OpenAI's master branch grew 1,378% between Q1 2024 and Q3 2026. The Q4 2026 estimate sits at 2,442%. Reliability fell below 90%. Fetches slowed, pushes stalled and merges failed. Pando was built in ten weeks, is fully compatible with the Git protocol, and has cut median push latency by nearly five times. It now serves close to a petabyte of Git data per day.

For enterprises, this is not a story about Git. It is a story about what happens when AI agents become so productive that the infrastructure underneath them simply stops working.

Why Pando exists

OpenAI's engineers were not doing 1,378% more work. Codex was. One engineer now has multiple changes ready for CI at once, and each push triggers several jobs that fetch the same commit. The compounding effect is machine-generated rather than human-generated. A single engineer with an AI coding agent produces the commit volume of several engineers working independently, and the CI system sees that volume multiplied across every automated test and build job.

The existing Git architecture was designed around human-scale work. Git grew out of a community that exchanged patches to the Linux kernel via email. For more than two decades, developers have adapted their tools around Git as the way they build software changed. But the shift from human commits to agent commits is not an incremental change in volume. It is a structural change in the nature of the workload.

OpenAI evaluated two conventional approaches before building Pando: a three-phase commit distributed system and a write-ahead log approach. Both proved too slow for the scale of concurrent pushes that Codex generates. Pando solves the problem by coordinating branch updates independently, so unrelated branches can be pushed in parallel, and by reusing packfiles across fetches instead of generating a unique packfile for every client. A client may download some objects it already has, but the server can serve the same bytes to many readers with minimal computation.

The result is a system that scales writes across branches and serves reads from caches without changing how engineers or agents use Git. That last point matters. Pando is a backend replacement, not a workflow replacement. Engineers and agents keep the same clients, the same commands and the same interface. Only the server underneath changes.

What this signals about agent-driven workloads

The significance of Pando goes well beyond source control. It is the first public example of a major technology company having to rebuild its own internal infrastructure specifically because AI agents exceeded the capacity of the systems designed for human work.

Most enterprise infrastructure was built on assumptions about human behaviour that do not hold when agents enter the picture. Consider a few examples.

Ticketing and case management. A human support agent handles one or two cases concurrently. An AI agent can work dozens simultaneously, each case generating its own subtasks, escalations, status updates and data queries. The CRM was designed for a person working through a queue, not for a system that processes queues at machine speed.

Document and contract workflows. A lawyer reviews a contract, makes redlines, forwards it to counterparts and waits. An AI agent can review, redline, compare against precedent, detect risk, generate a summary and route the document in minutes. The document management system was designed for human review cycles measured in hours or days.

Data pipelines. A data analyst writes a query, inspects results, refines the logic and reruns. An AI agent can generate queries, inspect results, detect anomalies, trigger downstream pipelines and alert stakeholders without pausing. The data warehouse was designed for bounded human workloads with predictable peaks.

In each case, the agent does not just do the work faster. It changes the fundamental shape of the workload. Human work tends to be bursty: periods of activity followed by natural pauses. Agent work tends to be continuous, parallel and cascading, where the output of one agent triggers the input of another across systems that were never designed to handle that throughput.

The infrastructure gap in enterprise AI

Enterprises are currently focused on proving that AI agents can complete tasks. They are testing agents for coding, document drafting, customer service, data analysis and process automation. Most pilots succeed at the task level. The agent generates the code, drafts the contract or resolves the ticket.

What the Pando story reveals is that success at the task level is only the beginning. The real bottleneck is whether the enterprise infrastructure can absorb the volume and pattern of work that agents produce when they succeed. A coding agent that writes ten functions a day is manageable. A coding agent that writes a hundred functions a day, each triggering CI builds, test runs, code review requests and deployment pipelines, is not manageable on the same infrastructure.

OpenAI is arguably the best-resourced AI company in the world. It employs some of the most capable infrastructure engineers on the planet. It had the luxury of observing the problem for months and building a custom solution in ten weeks. Most enterprises do not have that luxury. Their ticketing systems, document repositories, CI pipelines and data warehouses are third-party products with fixed scaling models and upgrade cycles measured in quarters or years.

The question for enterprise buyers is therefore not whether agents can do the work. It is whether the systems that consume the agent's output can keep up.

What enterprises should do now

Pando offers a useful framework for thinking about enterprise infrastructure readiness.

Model agent throughput, not just agent accuracy. Most enterprise pilots focus on measuring whether an agent completes a task correctly. Equally important is measuring what happens to the systems downstream. How many tickets does the agent generate per hour? How many document versions does it create? How many CI builds does it trigger? How many database queries does it execute? If the answer is "we have not measured that yet," the pilot has not tested scale.

Audit infrastructure for agent-shaped workloads. Many enterprise systems have rate limits, concurrency caps and throughput thresholds that were set years ago against human usage patterns. A ticketing system that caps at fifty concurrent users is not designed for an agent that generates fifty tickets in ten minutes. Enterprise buyers should work with their infrastructure and platform teams to identify which systems have hard limits and where those limits sit relative to projected agent workloads.

Build observability for agent activity, not just human activity. Most enterprise observability stacks were built to track human users: who logged in, what they accessed, how long they stayed. Agents do not behave like users. They may execute thousands of operations without a login session. They may generate requests from multiple identity contexts. They may create events that never appear in the UI because they happen through APIs. Observability needs to capture agent-generated load, not just human-generated load.

Prepare for infrastructure replacement, not just agent adoption. OpenAI's decision to replace its Git backend rather than patch it is instructive. At a certain point, the volume gap between human and agent workloads becomes too large for incremental optimisation. Enterprises should identify which systems are candidates for replacement versus augmentation, and budget accordingly.

What it means for suppliers

For companies building agentic AI platforms, Pando is both a warning and an opportunity.

The warning is that agent success creates its own problems. If your platform enables agents to work at scale, you need to be able to articulate what happens to the systems around them. Can your agent generate work faster than the customer's infrastructure can consume it? If so, what monitoring, throttling and backpressure mechanisms do you provide?

The opportunity is that infrastructure vendors who solve the agent-throughput problem will have a significant advantage. Just as Pando gave OpenAI a Git backend that could handle agent-scale commits, enterprises need ticketing systems, CI pipelines, document repositories and data warehouses that can handle agent-scale throughput. The vendors that build agent-aware infrastructure will differentiate themselves from those that simply bolt AI onto human-scale products.

The Agentic Expo angle

Agentic Expo 2027 at Olympia London on 23 to 24 March 2027 exists because the agentic AI shift is generating questions that no single vendor or platform can answer alone. Pando is a powerful reminder that the conversation needs to include infrastructure, not just models and agents.

The buyers at Agentic Expo are not just evaluating which agent platform to buy. They are asking whether their existing systems can survive the scale of work those agents will produce. The suppliers who can demonstrate not just agent capability but also infrastructure compatibility, throughput management and observability for agent workloads will be the ones that close deals.

By March 2027, the first wave of agent deployments will have moved from pilot to production. The organisations that are preparing their infrastructure now, rather than discovering its limits after agents go live, will be the ones that capture the operational advantage.

Get Tickets Exhibit at Agentic Expo

Sources: OpenAI, "Scaling Git at OpenAI", 2 November 2026; Interesting Engineering, "OpenAI builds custom Git backend in 10 weeks after 2,442% commit surge", 3 November 2026; FourWeekMBA, "OpenAI's Pando Explained: A Git Backend Serving ~1 PB a Day", 3 November 2026. Source code: github.com/openai/git.