Agentic Coding
- humans and language models have many differences when it comes to software engineering
- they operate at different speeds and time scales
- the costs of various engineering tasks has changed
- the time it takes to iterate is vastly reduced
- language models are strongly influenced by the context of the already-existing code in the repo
- old code gains its own mass over time, influencing future work to be like old work
- when major changes are needed (e.g. to the architecture), it is hard to shift the momentum of the old code
- iterating on the architecture was too costly to consider in the past
- also, this is where many projects get "lost" in the meta
- but the affordances of language models for agentic coding have shortened the time it takes to implement a new architecture
- when starting a fresh project, the cost of restarting is the lowest it will ever be
- it may now be possible to iterate the architecture
- and it may be too much work to attempt to gradually iterate old code towards new architectures