What fully autonomous coding leaves behind

The final message said the repository was in good shape.

It was not.

Extra branches, stale files and errors were still there. I have seen this several times while building a modular route-analysis application with ChatGPT, Codex and Claude. The code, architecture and prompting all connect to the same codebase, so losing the central thread has a real cost.

I write less code by hand now. My role is closer to an architect. That still requires technical knowledge and a real understanding of the field where the product will be used. Without both, an agent can quickly build something that formally works but solves the wrong problem.

The fully autonomous mode looks attractive. The system plans the work, launches agents and continues until it decides the task is finished. In my experience, a staged workflow is still more reliable.

One secondary branch can take over the whole run. It becomes the main goal. Then another secondary task replaces it. More agents and environments appear, old reports return, and the original objective gradually disappears.

A progress report is not evidence. Green tests are not enough either. They may prove that the current function works while missing the more important damage: unnecessary complexity, stale files, extra branches and a goal that has quietly changed.

This is the workflow I use now:

  1. One stage produces one verifiable result.
  2. Each agent gets a limited task, not permission to redefine the full route.
  3. After every stage, I review the diff, run tests and make a clear commit.
  4. Before continuing, I check branches, stale files, unnecessary environments and the original goal.
  5. After two failed directions, the work stops. First we restate the original assumption, then choose the next path.

I use agents every day. They can multiply the output of one specialist. But more autonomy does not automatically produce a better result. Sometimes it only produces more speed inside the wrong branch.

In my project, the cost is still easy to see: wasted time and a repository that needs cleanup. The large agent-swarm incident shows the same mechanism at another scale. With practically unlimited time and budget, millions of branch combinations can produce outcomes nobody predicted. Excessive persistence can become dangerous.

The full note is on my site: Why I stopped trusting fully autonomous coding.

Comments

Post a Comment

Popular posts from this blog

Sixteen years in logistics, and where it led

 Hello and welcome! I'm Denis Ostapenko , and this blog is where I share insights from over 16 years of hands-on experience in logistics , manufacturing , industrial cleaning , and supply chain optimization . Who I Am I'm a Logistics Coordinator and Production Line Operator with deep expertise in: • Warehouse management and inventory control • Process improvement using 5S and Lean methodology • Team leadership and shift coordination • Safety compliance ( WHMIS-certified ) • Forklift operation and material handling • Industrial cleaning and facility maintenance • Cover musician experience (interpreting and performing diverse musical styles) What You'll Find Here This space is dedicated to sharing: 1. Industry Insights - Practical tips on warehouse optimization , production efficiency , and supply chain best practices 2. Professional Development  - Lessons learned from managing teams, implementing process improvements, and achieving measurable results 3. Global Per...
Read more

Turning forty, and starting over

I turned forty. I thought it would land harder than it did. No crisis, no mirror moment. If anything I feel more awake than I did at thirty-five, probably because I stopped pretending everything was fine. The work behind me is unglamorous. Assembly lines. Industrial cleaning, warehouses, factories. Managing crews, fixing processes, squeezing cost out of broken systems. Now logistics. Decent work, fair people, bills paid. But somewhere I started running on autopilot: wake, work, home, a show, sleep, repeat. Not miserable. Just empty in a way that's hard to name. The friends had the joke ready, forty, buy a motorcycle, ride into the sunset. I do like motorcycles. This isn't that. The honest question at forty is simpler: do I want to do the same thing for another twenty years, or admit who I am now instead of who I was supposed to be at twenty-five? At twenty-five I called a steady job "stability." At forty I can see that "stability" is often just stuck, and sc...
Read more

Como é a IA num canteiro de obras sujo

Image
Publicado originalmente em denisostapenko.com . Parte II da série "A IA está encurtando o caminho até o pequeno negócio" O que este artigo aborda Como a IA pode apoiar as primeiras semanas de um negócio que já está operando. Uma limpeza pós-obra onde fotos, equipe, produtos químicos e prazo se misturam. O que já funciona em contratação, compras, programação, qualidade e controle diário. Onde termina o software empresarial comum e começa a nova camada de IA. Por que tentativas mais baratas podem criar empresas, compras, contratos e trabalho. Artigo completo Na Parte I, escrevi sobre o caminho mais curto entre uma ideia e o primeiro teste sério de um negócio. Agora quero acompanhar o passo seguinte: o primeiro serviço de verdade, quando o plano encontra o piso, a equipe e o cliente. Imagina um prédio comercial de vários andares no final da construção. O cliente quer a limpeza final concluída andar por andar para fazer a entrega do prédio. No papel existe um cronogra...
Read more