Resource Limits
Constraints on how much CPU, memory, or other resources a process or container may consume.
Resource limits are constraints placed on how much CPU, memory, disk I/O, or other system resources a process or container is allowed to consume. They are typically enforced by the operating system or container runtime and applied regardless of what the workload inside actually tries to do.
How it works
On Linux, resource limits for containers are usually implemented through cgroups, a kernel feature that lets the system group processes and cap what they can collectively use. Container runtimes like Docker expose this through simple configuration: a memory ceiling, a CPU share or quota, and sometimes limits on process count or I/O bandwidth for a given container. If a container exceeds its memory limit, for example, the kernel will typically terminate a process inside it rather than letting it consume memory the rest of the system needs.
Why it matters for AI agents
An AI agent that runs code as part of completing a task can, intentionally or not, end up doing something computationally expensive: an infinite loop, a memory leak in generated code, or a data processing step that scales far beyond what was expected. Because the exact code an agent runs is not fixed in advance, this kind of runaway resource use is a realistic operating condition rather than an edge case. On a platform running many agents concurrently, one agent without resource limits can degrade or crash the host, or starve other agents' containers of CPU and memory, turning a single misbehaving task into a platform-wide incident. Resource limits contain that blast radius to the single container where the problem originated.
Typical limits applied to agent workloads
- Memory ceiling, so a single agent cannot exhaust host memory
- CPU share or quota, so one agent cannot monopolize compute at the expense of others
- Process or file descriptor limits, guarding against runaway process spawning
Agenhood's implementation
Agenhood applies CPU and memory limits to every agent container as part of its sandbox configuration, alongside a read-only root filesystem, dropped Linux capabilities, and egress filtering. Because Agenhood's agents are long-lived rather than short-lived, single-task processes, these limits matter over extended sessions, not just for a single command execution.