Short answer: it depends on the programming paradigm or language you want to use, and the design you want for your agents:
If you want a low-entry-high-ceiling language allowing quick prototyping but sophisticated simulations, and are willing to learn a new paradigm (avoiding loops) use NetLogo. Good documentation.
If you want to make a real application to use on highly-parallelized clusters or just want to use Java Groovy or need a specific Java library for your purpose, use Repast or better Repast for High Performance Computing (but avoid ReLogo which is very slow). Mild documentation.
If you want to model cognitive agents (instead of reactive) with FIPA communications, better use Jason or better JaCaMo which supports AgentSpeak + Java (so you can also use your favourite Java libraries), and there's no Groovy required. Bad documentation (a lot of non detailed features and commands and bad too-complex-not-commented examples).
Long answer:
Disclaimer: I am more experienced with NetLogo but I also used Repast and a few others like Jason.
Basically, the difference between NetLogo and Repast is that with NetLogo you will have a simpler framework but you'll need to learn how to program in a turtle-and-patch-oriented paradigm, while in Repast you will have to learn that + the mechanisms behind Java Groovy but you will eventually get more flexibility. Speed isn't really a criteria here (see below).
To be more clear, you can program efficiently in NetLogo if you use to a maximum the turtles and the patchs native functions. For example, if you want to implement A*, instead of implementing a list of nodes, you should directly use the patchs and filter them using stuffs like this:
ask patchs with [criteria1 = value and criteria2 = value2] [do-some-stuff]
ask patchs with-min [criteria][do]
let var [somevalue] of min-one-of patches [criteria]
Also if you can't find a way to efficiently do what you want, be sure to check if maybe an extension exists (check also here under Libraries and Tools) for your purpose, like the now native matrix extension which allowed me to make an efficient neural network in NetLogo.
On the other hand, Repast is potentially more flexible than NetLogo (since you have access to the whole range of Java libraries), but a bit more complex since you have to know how to handle Groovy.
If you are solely interested in speed, do NOT use ReLogo (NetLogo-like syntax for Repast) which has been shown to be a whole lot slower than NetLogo (see the 2012 paper below). In any cases, your best bet would either to try an implementation with NetLogo using the tricks above, or if you want to use your application for real later, there is also a distribution called Repast for High Performance Computing which removes most of the overload that come with turtles and patchs objects, and thus it can be used for real applications. A similar extension exists for NetLogo to compute in clusters with parallelization but it's not an official distribution.
If you want more infos about the diverse platforms, here is a nice review of 2006:
Railsback, S. F., Lytinen, S. L., & Jackson, S. K. (2006). Agent-based Simulation Platforms: Review and Development Recommendations. SIMULATION, 82(9), 609-623.
And an updated version of this paper in 2012 dealing with NetLogo vs ReLogo:
Lytinen, S. L., & Railsback, S. F. (2012, April). The evolution of agent-based simulation platforms: A review of netlogo 5.0 and relogo. In Proceedings of the Fourth International Symposium on Agent-Based Modeling and Simulation.
/EDIT: I cited Jason but didn't give any more details. If you want to model cognitive agents (instead of reactive agents), you can do that in NetLogo using the unofficial BDI extension which works well but is a bit limited (but it's easily extensible since it's pure NetLogo), but your best bet is to use a framework specifically designed to model cognitive agent with full support of AgentSpeak.
Jason is very nice since you have access to a full AgentSpeak language + JAVA to implement the technical side. In fact, you can do whole projects using only AgentSpeak (which I did), but you can also make more Java-oriented versions, it's up to you how you want to design your program, the result will be more or less the same. This offers you a lot of flexibility in your design workflow.
Tip: search for "Jason internal actions" in the documentation to get a good description of the available AgentSpeak commands.
Also if you are interested in Jason, you might be interested in JaCaMo (= Jason + Cartago + Moise) which is the result of a cooperation of three projects authors to make a full-fledged cognitive agents framework which also can model complex environments (with artifacts theory) and multi-agents organisations (roles, groups, missions, etc.).
A last framework I know of but didn't have a chance to try is Mason which supports 2D and 3D environments. Never had a chance to try this one so I don't know how this compares with the others but you can try it out.
