Agents working at the same time
Agent concurrency is the maximum number of agents which can be working at the same time.
For most software development processes, having just one agent running at a time might not be the quickest way to get work done. Some work can be parallelized - for instance, if there are multiple sections of the code base, or if one agent is doing an implementation and another agent is doing research.
Concurrency becomes important when the work starts to overlap - for instance, both agents are working on the same source code file, or working on the same function or class within a file. Or both agents are modifying a test suite to cover their implementation. Other areas start to show up when work gets integrated, and some work was done on a stale main branch that now needs to be updated.
Total agent count is just one part of the strategy
Agents are intelligent, and you can use that to improve the throughput of your software development process.
For instance, some agents might take on more than one task in a session. Other agents might coordinate with each other to work on different parts of the code base, or even to acquire an exclusive lock on part of the code base while the agent completes its work to avoid rework.
It’s important to consider what counts as complete when doing an analysis of agent concurrency - it typically isn’t just when one agent is finished with a task - it’s once the work is merged into main and the tests pass.
Learn more at InstantRocket
The Agent Concurrency Lab lets you vary the number of slots, shared resources, and integration costs in a simplified model. Experiment 001 reports measurements from a real coding workload, with its method and limitations. You can also take the harness from Experiment 001 and extend it to your own project or coding harnesses, and generate your own results.