Plan to Throw One Away
Chapter 11 of The Mythical Man-Month bears the title 'Plan to Throw One Away.' Brooks's argument was that the first implementation of any non-trivial system is not the final product but a learning experiment. The builder cannot know, before building, what the system should actually do. She learns by building a version that approximately does what she thinks it should do, discovering through the implementation the gap between her initial understanding and the problem's actual shape. The prescription follows: plan for this from the start. Build the first version knowing you will throw it away. Build the second version with the understanding the first produced. AI has made the economics of this prescription dramatically more favorable — the first version can be built in hours, not months — but the learning it enables still depends on the builder's capacity to evaluate what the first version reveals.
In the AI Story
Brooks's insight was empirical. He had watched the OS/360 project and other large IBM efforts, and he had noticed that teams which treated the first implementation as the final product consistently delivered worse systems than teams which treated the first implementation as exploration. The difference was psychological: the team that planned to throw the first version away was free to experiment, to take risks, to discover what would not work. The team that expected the first version to ship was conservative, risk-averse, and blind to the gap between their initial specification and the actual problem.
The prescription was controversial in management contexts because it seemed to license wasted effort. Managers wanted the first version to be the shipping version; they did not want to pay for two implementations when they thought they were paying for one. Brooks's counter-argument was that they were going to pay for two implementations regardless — because the first version, forced into production, would be bad enough that a rewrite would be necessary within a few years. Planning for the rewrite from the start produced a better second version at lower total cost.
The Orange Pill moment has transformed the economics of this prescription. The first version that used to take six months now takes six hours. The cost of throwing it away has collapsed proportionally. In principle, this should be an unalloyed improvement: cheaper exploration means more exploration means better final versions. In practice, the improvement is partial.
The Brooks volume argues that the speed of generation does not reduce the difficulty of evaluation. The builder can produce ten prototypes in a week, but she must still examine each prototype, understand what it reveals, and integrate the learning into her understanding of the problem. The cognitive work of learning from the prototype has not compressed as dramatically as the cognitive work of producing it. The result is that many AI-era builders produce more prototypes without extracting proportionally more learning — and therefore without developing the understanding that the throw-one-away prescription was designed to produce.
Origin
Brooks's prescription appeared in Chapter 11 of The Mythical Man-Month, titled 'Plan to Throw One Away: You Will Anyway.' The formulation was deliberately provocative; Brooks wanted to force managers to confront the empirical reality their budgets and schedules denied.
The anniversary edition's The Mythical Man-Month After 20 Years chapter revised the position slightly. Brooks acknowledged that the prescription had been misunderstood: he was not recommending that teams always build two versions, but that they build the first version with the honest expectation that significant parts of it would not survive into production. The 'incremental build' and 'evolutionary development' methodologies of the 1990s were consistent with this refined position.
Key Ideas
Exploration as purpose. The first version's purpose is to discover what the system should be, not to be what the system is.
Honest economics. Managers who expect the first version to ship pay for two versions anyway; the prescription is to pay openly rather than through crisis.
The learning requirement. Throwing one away requires extracting the learning; the speed of generation does not automate the understanding.
AI compression. The cost of the first version has collapsed under AI; the value of the prescription depends on whether the builder's evaluative capacity has kept pace.
Debates & Critiques
The relationship between Brooks's prescription and the iterative methodologies that came after it is contested. Some read agile development as a vindication of Brooks — building in small increments with the expectation that each increment will be revised. Others read it as a modification — Brooks wanted a full first version thrown away, while agile prescribes continuous partial revision. The AI era adds a third possibility: generation is now so cheap that the distinction between versions collapses into continuous regeneration, and the prescription's original logic requires reformulation.
Appears in the Orange Pill Cycle
Further reading
- Frederick P. Brooks, The Mythical Man-Month, Chapter 11 (Addison-Wesley, 1975)
- Eric Ries, The Lean Startup (Crown, 2011)
- Barry Boehm, A Spiral Model of Software Development and Enhancement, Computer 21, no. 5 (May 1988)