By 2018, it was easy to believe the cloud wars were over. It seemed as though the big providers had won, and there was little room for anyone else. Meanwhile, developers were still spending weeks getting applications ready to run, and the industry had come to accept how much work the cloud required, even when the application itself was relatively simple.
Today, developers are building applications that have been sitting in their backlogs for years, and with the help of coding agents, they’re doing it in days rather than weeks or months. But all of this new software needs somewhere to run, and the agents writing it are now being asked to deploy it and set up the infrastructure it needs.
For coding agents, deployment means moving beyond the repository and into the infrastructure layer, where the consequences of a mistake are much higher. But to work on infrastructure safely, they need more than an API key and permission to deploy.
How Cloud Infrastructure Must Adapt for Coding Agents
To allow coding agents to safely deploy software end-to-end, cloud platforms need to evolve beyond code repositories into the infrastructure layer by providing:
- Full System Context: Access to databases, APIs, security policies and environment variables.
- Predictable Feedback Loops: Machine-readable outputs, logs and clear failure diagnoses.
- Safe Operations: Idempotent retries to prevent duplicate resource creation.
- Scoped Permissions and Approvals: Task-specific access controls with human approval checkpoints for high-risk changes.
- Concurrency and Spend Limits: Protections against machine-speed activity exceeding budget or API constraints.
Agents Need to Work Across the Whole System
Developers keep applications on existing infrastructure for good reasons. Security policies may limit where software can run, and a new service often needs to connect with databases, internal APIs and other systems already on the company’s network.
Many cloud environments, however, are difficult for an agent to operate on as a whole. The information it needs may be spread across several services, dashboards and internal documents. An agent might be able to deploy the code but lack the context or access required to connect it safely with the rest of the system. That ultimately forces a handoff back to a person, creating more work in the long run.
To run a fully agentic loop, an agent needs to understand the full context of the environment it’s working in: which resources it can use, how to handle environment variables and security, which adjacent systems might be impacted by a change, how to run a full deployment from previewing changes to rolling them into production. Engineers need full visibility and control to set agent permissions, review changes and step in when necessary.
Infrastructure Has to Be Legible to Agents
Coding agents are effective within a repository because they can inspect the entire codebase, make changes, run tests and see whether they worked. Infrastructure needs to offer a similar feedback loop.
Giving an agent permission to push code isn’t enough. It needs to understand the state of the running application before it acts, follow the new deployment as it runs and verify the outcome afterward. When something fails, it needs enough information to diagnose the problem without having to hand the task back to an engineer.
A cloud platform built for this way of working should provide all of the following.
Structured, Predictable Responses
Agents handle clear statuses and machine-readable output better than vague messages or information buried in a dashboard. They need to know which resources exist, what changed and why an operation failed.
Safe Retries
If an agent loses its connection halfway through creating a database, repeating the request shouldn’t create a second database. Cloud operations need to account for agents retrying tasks after an incomplete response.
Scoped Permissions
An agent might be allowed to create a temporary test environment but prevented from deleting production data. Permissions should reflect the task instead of giving the agent broad access to an account.
Approval Points for Consequential Changes
An agent can inspect usage and recommend more compute without being allowed to increase the bill on its own. Teams should be able to decide which actions require a person to approve them.
Access to the Full Operational Loop
Starting a deployment is useful. Reading its logs, checking its health, rolling it back and cleaning up temporary resources allow the agent to finish the job.
Consider a pull request that changes an application and its database. An agent could create an isolated environment, provision a temporary database, deploy the branch and run integration tests against it. Once the pull request is merged, it could remove the environment. The engineer remains responsible for setting the rules, while the agent handles the repeatable work inside them.
This type of loop enables agents to do real, autonomous work safely, but then it also needs to scale. Coding agents can work across several branches, environments and applications at once. The infrastructure layer has to account for a pace and volume of activity that its original limits may not have anticipated.
Cloud Limits Were Built for Humans
A developer typically creates a service, waits for it to deploy and then decides what to do next. An agent can create several environments in parallel, inspect each deployment and retry failed operations within seconds.
That behavior can easily exceed limits that seemed reasonable when a person was on the other end. It can also turn a small mistake into a larger one. A retry loop might create duplicate resources. Several agents could compete for the same database connections or continue provisioning compute after a team’s intended budget has been reached.
Cloud providers will need to distinguish between harmful activity and legitimate work happening at machine speed. Higher limits alone won’t be enough. The underlying operations need protection against duplicates, along with controls over concurrency and spending.
Teams will also need a clear view of what their agents are doing while the work is happening. A monthly bill or an audit log reviewed after an incident comes too late. Engineers need to see which agent created a resource, why it created it and whether that resource is still needed.
The same qualities that make coding agents productive can make them difficult infrastructure users. They move quickly, repeat tasks and work in parallel. Cloud platforms built for agents should actively enable them to operate as quickly as possible without compromising the reliability or security of the system.
What Will the Next Cloud War Winners Win?
Cloud platforms will be judged by whether an agent can complete a real task, recover when it fails and leave behind a record an engineer can trust. Engineers need to be able to see what the agent did, limit what it can change and step in when the risk calls for it.
The same platform has to support two ways of working. People need room to make decisions and investigate problems. Agents need operations they can execute consistently, with clear responses they can act on. Designing for one while neglecting the other will leave part of the development process broken.
Existing contracts, policies and infrastructure will continue to influence where applications run. But as agents take on more of the work, familiarity alone will carry less weight. The next cloud leaders will be the providers that both developers and their agents can trust to get software into production and keep it there.
