Quick answer
Coding agents have moved the developer's job up a level. The skills that gain value: writing precise specifications and acceptance criteria, reviewing large volumes of generated code quickly, designing systems and interfaces, testing strategy, operating agents (permissions, sandboxes, costs), and understanding the codebase deeply enough to know when the agent is wrong. Raw typing speed and memorising APIs matter less. Fundamentals matter more, not less.
The fear is that agents replace developers. The reality on teams that use them heavily is that they replace a specific part of the job, the part between "we know what to build" and "it is built", and expand everything on either side of it.
Specify
Agents are only as good as their instructions. Writing requirements with acceptance criteria, describing interfaces precisely, and maintaining project rules files (the steering documents Kiro, Cursor, and Claude Code all read) has become a core skill. Developers who can turn a vague request into a spec an agent can execute are the ones teams now value most.
Review
When an agent produces a 400-line pull request in ten minutes, the bottleneck is reading it. Effective review means knowing what agents get wrong (edge cases, error handling, security, invented APIs), reading diffs strategically, and using a second agent for the first pass. AI code review tools help; they do not replace the human who understands the system.
Design and fundamentals
Data modelling, API design, concurrency, security, performance: agents implement these well only when told what to do, and they hide mistakes behind working code. The developers who catch the wrong data model before it ships are the ones with fundamentals. Learn them properly; the agent will not.
Operate
- Permissions and sandboxes: what an agent may read, write, run, and access
- Cost: an agent session on a frontier model costs real money; know the levers (model choice, context size, caching)
- Tooling: MCP servers, hooks, and CI integration for agent workflows
- Evaluation: knowing whether a new model or prompt made the team faster or just noisier
A practical plan
- Use an agent daily on real work for a month; notice where it fails
- Write a spec before every non-trivial task and see how much the output improves
- Read every line an agent produces for a week, then decide what you can safely skim
- Pick one fundamentals gap (testing, security, data modelling) and close it
Related reading
Bottom line
The job is not going away; it is moving toward the parts that were always the hard parts. Learn to specify, review, design, and operate, and the agents make you more valuable, not less.
