Engineer abandons manual coding for six months using AI agents exclusively

The experiment demonstrates a shift from manual coding to agent-driven development, where the engineer’s role became one of specification and validation rather than direct implementation. The AI agents generated code across multiple files, but the engineer still had to review every change to ensure correctness. This mirrors traditional code review but with a higher error rate, as agents frequently introduced unnecessary or incorrect edits. The trade-off was clear: less typing, but more cognitive load to steer the agents toward the desired outcome.

Parallel agent workflows emerged as a natural consequence of the engineer’s reduced manual involvement. While spinning up multiple agents filled idle time, it created new coordination problems. Agents competed for resources like ports, databases, and Git state, leading to conflicts that required manual intervention. Workarounds like worktrees and containers helped but did not fully resolve the issue, as agents could still interfere with shared system resources. This suggests that agent-based development may not scale efficiently without better isolation mechanisms.

The experiment also revealed the limitations of current agent tooling. Agents struggled with context management, often wasting tokens on conflict-avoidance instructions rather than task execution. The engineer’s solution, iterative prompting and environment tweaks, required hands-on debugging, much like traditional software development. This implies that agent-driven workflows may not yet reduce the total effort required for complex tasks, but they could shift the nature of that effort from implementation to oversight.

A key insight was the engineer’s reliance on a mental model of the codebase to guide the agents. This model, built over years of manual work, was critical for validating agent output and identifying gaps in their understanding. Without this deep contextual knowledge, the agents’ errors would have been harder to catch. This raises questions about whether less experienced engineers could replicate the experiment’s success, or if agent-driven development will remain dependent on human expertise for the foreseeable future.