What does software development look like when agents write 100% of the code?

Summary

The post argues that software development should shift to isolated, long-running coding agents that write production code under human-written specs and verification. Engineers spend less time editing generated code and more time defining work, context, tests, and release checks.

Problem

Parallel coding agents can interfere with a developer’s local machine through shared ports, processes, state, dependencies, and shell access.
Long-running agent tasks need compute that stays online and remains isolated when the developer is offline.
Human micromanagement of many agent sessions becomes the bottleneck, so output quality depends on better upfront specs and verification.

Approach

Run each coding agent on its own isolated computer, separate from the developer’s machine.
Give agents detailed product specs, architecture decisions, and task breakdowns before they start coding.
Close the feedback loop with unit tests, integration tests, end-to-end tests, linting, type checks, observability, and task-specific tools such as agent-browser and agentmail.
Move human review toward specs, context, and verification processes, with PR review as a final check before merge.

Results

The post gives no benchmark, dataset, controlled experiment, or measured accuracy result.
It claims agentic coding moved in 2 years from toy autocomplete to generating an overwhelming share of production code, but it gives no percentage or study.
It claims implementation time can shrink from weeks to hours when agents have clear specs and closed-loop verification.
It claims isolated agent computers reduce collisions across ports, processes, local state, and dependencies during parallel work.
It claims red/green TDD and comprehensive tests let agents iterate toward a PR with less human steering.

Link

https://blog.bastion.computer/what-does-software-development-look-like-when-agents-write-100-of-the-code/