Legacy code can mean a number of related things, the most common of which are code inherited from someone else; code from a previous version of the application; code that is deprecated or outdated; or code that is untested. Many legacy systems meet all these criteria, and legacy code typically requires a different, more delicate approach than new code.
From the Wikipedia entry (note that the article has been flagged for lacking citations):
Legacy code is source code that relates to a no-longer supported or manufactured operating system or other computer technology. The term can also mean code inserted into modern software for the purpose of maintaining an older or previously supported feature — for example supporting a serial interface even though many modern systems do not have a serial port. It may also be in the form of supporting older file formats that may have been encoding in non-ASCII characters, such as EBCDIC.
In practice, most source code has some dependency on the platform for which it is designed—unless a programmer uses a platform-independent programming language like Java, it is hard to write a large, useful program that is totally independent of its environment. When the manufacturer upgrades a platform (or the platform is superseded), the code may no longer work without changes, and becomes legacy code. A large part of the task of a software engineer is to continually alter code to prevent this.
While the term usually refers to source code, it can also apply to executable code that no longer runs on a later version of a system, or requires a compatibility layer to do so. An example would be a classic Macintosh application which will not run natively on Mac OS X, but runs inside the Classic environment, or a Win16 application running on Windows XP using the Windows on Windows feature in XP.
More recently, the software engineering community has developed other interpretations for the term legacy code. Among the most prevalent are source code inherited from someone else and source code inherited from an older version of the software. Michael Feathers' Working Effectively with Legacy Code (ISBN 0-13-117705-2) introduced a definition of legacy code as code without tests, which reflects the perspective of legacy code being difficult to work with in part due to a lack of automated regression tests. He also defined Characterization Tests to start putting legacy code under test.