Datadog Agent

What the Agent collects and what it sends

The Datadog Agent is a resident program that keeps running on the monitored host. It periodically collects host metrics and events such as CPU, memory, and disk, and sends them to Datadog. With additional configuration, the same Agent also handles log collection, receiving traces for APM, and gathering information about running processes. Rather than installing a separate collector for each kind of data, the structure is one Agent per machine that serves as the single point of entry. As a reference, the official documentation lists a measurement in which Agent 7.34.0 running on an AWS c5.xlarge instance averaged roughly 0.08% CPU usage and about 130 MB of memory (RSS) (as of August 2026; source: Datadog Agent documentation). The load on the monitored host is on the fairly light side, but the same page also records a measurement in which enabling log collection on the older Agent 6.15.0 raised CPU load (measured under different conditions from the 7.34.0 figure above), so it cannot be treated as exactly zero.

Installation basics - from major operating systems to containers

Installation procedures are provided per environment. The official documentation covers Linux distributions including Amazon Linux, Windows, and macOS, as well as container environments such as Docker, Kubernetes, and Amazon ECS, and installation methods for Heroku and AWS Lambda (as of August 2026). For a handful of servers, running the distributed install command on each machine is enough, but as the count grows, distributing the Agent through the official recipes for configuration management tools such as Ansible, Chef, Puppet, and SaltStack is the realistic approach. There is also a management feature called Fleet Automation that lets you perform post-installation upgrades and configuration changes in bulk from the browser, and Datadog officially recommends it for managing installation and updates in large environments.

The configuration file and the status command

The central configuration file of the Agent is datadog.yaml, where you write basic settings such as site, which determines the region data is sent to. The first thing to memorize right after installation is the status command.

sudo datadog-agent status

The Running Checks section of the output lists, for each enabled integration, the number of runs, the number of metrics collected, and the average execution time. If no warnings or errors appear there, that check is working correctly. Starting and stopping the Agent as a service follows the conventions of each environment: on Linux it is service datadog-agent start / stop, while on Kubernetes deleting the Agent's Pod causes it to be rescheduled automatically (source: Agent Commands).

When to use agentless integrations instead

Datadog provides integrations with cloud providers, so AWS, Google Cloud, Azure, vSphere, and others can be monitored without installing the Agent by ingesting the measurements exposed through the cloud provider's API. So when does the Agent become mandatory? The dividing line is whether the information is visible from outside the host or only from inside it. What a cloud API returns stops at values observable from the outside, such as per-instance utilization; listing processes inside the OS, application traces, and collecting log files on the host are the Agent's job. In practice, a two-tier arrangement is easy to build: first cover the whole environment thinly with cloud integrations, then install the Agent only on the hosts you want to see in depth. Note, however, that VMs ingested agentlessly are also counted as billable hosts, as described in the next section, so keep that in mind as a premise of your estimate.

The relationship to host billing - it is not decided by the Agent count alone

The Agent as a piece of software has no corresponding line item on the official pricing table as of August 2026. Charges arise on the product side, such as Infrastructure Monitoring, and the unit is the number of hosts. What is easy to misunderstand here is the counting scope: the official billing documentation states explicitly that, in addition to instances with the Agent installed, VMs monitored through the AWS, Google Cloud, Azure, and vSphere integrations are also counted as hosts (a VM that has the Agent is not double-counted). The count is taken hourly, and at the end of the month the maximum value after excluding the top 1% of that month's measurements (the high-water mark) becomes the basis of the invoice. Short-lived increases from autoscaling have little effect on the bill, whereas a permanent increase in host count partway through the month sets the level of that month's invoice (source: Datadog billing documentation).

A detailed reading of the pricing is compiled in our guide to the pricing structure.

Version management and update pitfalls

As of August 2026, the current major version is Agent 7, and Datadog officially recommends updating with each minor and patch release, at least once a month. Monitoring software is an area where "it works, so don't touch it" easily takes hold, but if you leave updates for a long time you fall short of the version required by a new feature or integration you want to use, and end up forced into one large leap of an upgrade at the worst moment. On the other hand, updating every machine at once means losing monitoring itself if the Agent has a defect, so a staged rollout is safer: update some hosts first, watch how they behave, then extend to the rest. In environments with many hosts, the shortest path around the pitfalls is to use Fleet Automation, which supports remote upgrades, or to pin the version in the configuration management tool to an explicit number rather than "latest" and treat updates as deliberate work.

Cleanup when stopping or removing the Agent

Stopping the Agent on Linux is service datadog-agent stop, and uninstall procedures are provided in the official documentation for each OS. What is worth knowing in relation to billing is the official statement that a host that has stopped sending data (shown as INACTIVE in the Infrastructure List) is removed from billing and can take up to about 2 hours to disappear from the list. However, as described in the previous section, the invoice is decided by the month's high-water mark, so removing the Agent partway through the month does not necessarily lower that month's bill right away. If you temporarily run the Agent on many hosts for testing, remembering the removal timing together with the billing measurement method will spare you a surprise when you read next month's invoice.