Win shares is also the name of the metric James describes in the book.
It considers statistics for players and boils them down to a single metric, a "win share", including all pitching, hitting and defensive contributions by the player. Statistics are adjusted for park, league and era. These points in themselves are not novel, but James' specific formulae are interesting.
The main innovation, however, is that a "win share" represents exactly one-third of a team win. Therefore, the aggregate win shares for all players on a team in a given year will sum to exactly three times the number of wins their team amassed. So if a team won 80 games in a season, then its players will share 240 win shares among themselves. (A player cannot be awarded negative win shares, though they may have zero.) This differs from other sabermetric player rating metrics such as Total player rating and VORP in that it is based on team wins, not runs.
One criticism of this metric is that it does not allow for the possibility of a team winning fewer or more games than the sum of its players' contributions suggests, which may happen due to unusual circumstances (which arise largely through chance and don't tend to even out during a single season). In short, it treats the arising and taking advantage of (or failing to take advantage of) circumstance as a component of player contribution. This blurs the line between whether win shares are intended to represent player value (what actually happened) or player ability (what the player was directly responsible for).
The Bill James Historical Baseball Abstract, 2001 edition, also written by James, uses win shares to evaluate the careers of many players, and to place them in contexts where they can be compared. The two books are effectively companions to one another.
Win Shares is very thoroughly critiqued in three articles by tangotiger, available at www.geocities.com/tmasc/