I have a swarm of agents, yet I must still be an expert. More so than ever before. There’s a planning agent, test-driven agent, coding agent, refactoring agent, documentation agent.

Yet none of them are fully trustworthy despite their overconfidence. It’s perverse that I’m ‘levelling up’ my analysis and development skills, of 20 years, like never before.

Needing to become a deep expert in all areas of the SDLC, just to ensure the swarm doesn’t become unwieldy. I’m now starting to perform the duties of Business Analyst and QA, out of necessity and likely to their detriment.

The upfront planning agent looked like magic to begin with, yet its inherent flaw soon became apparent. It simply cannot decide which architectural patterns are best, worse, or indifferent. The transition from a well-structured codebase and interfaces to a dog’s breakfast is only a few autonomous commits away.

A few rounds of conversation with the planning agent will iron it out, but I’m still the one deciding, and that’s informed by years of handcoding C#. Same for overseeing code changes, test coverage, PR reviews. But what would happen if an entry-level and junior developer faced my swarm? It wouldn’t be pretty.

Vibe coders say it doesn’t matter, as long as the tests are green and functionality works as expected. But how are technical debt, ease of change, and resilience being factored in? I’m not sure they are.

Agentic coding is certainly fascinating, but the opportunity for things to go very badly wrong seems greater now than ever before. Guardrails and checks and balances will be top of mind today, as I engage with the swarm for another round.

Get an email whenever Frank Ray publishes. Subscribe here