Не пишу код уже 2 недели. Нужны ли еще разработчики?

Summary

The speaker describes a work project developed using Spec-Driven Development (SDD). Instead of writing code manually, the team gives an AI a Jira ticket and supporting specifications, then has it generate and modify the code.

Their setup uses OpenCode as the agent harness, OpenSpec to manage requirements and plans, and a corporate GLM model. They are also adding agents and skills for tools such as Jira and Confluence.

The workflow is roughly:

1. Give the AI agent a ticket.
2. Review its proposal and implementation plan.
3. Apply the plan and let the agent generate the code.
4. Check that the result works and address any remaining issues.

The speaker found that closely reviewing every line of the AI’s initial plan was unproductive. Their role has shifted toward describing the task, checking the outcome, and guiding the agent. One large generation task took about three hours. In another example, generating controllers from Confluence contracts took around 20 minutes, compared with what the speaker estimates would have been a day of manual work.

Key observations and concerns

• The work is changing more than disappearing. People are still needed to configure agents, provide context, improve specifications, and judge whether generated results are correct.
• Programming-language expertise may matter less in this workflow. The speaker felt that the agent’s ability to work across languages made their accumulated knowledge of particular languages less central. They question whether traditional interview topics, such as language internals, will remain as important.
• Generated code may be difficult to treat as a reliable black box. The speaker expects testing, security review, and load testing to remain important. Automated debugging and testing are promising, but their own attempts were slow.
• AI can help with merges. The speaker describes asking an agent to merge the latest master branch and resolve conflicts, including subtle discrepancies in generated specifications. They found this useful but still reviewed the result.
• Requirements and analyst involvement may change. If requirements are clear, an agent could potentially ask an analyst questions directly. For now, the speaker says setting up and configuring agents requires technical skills that analysts and testers may not have.
• The profession’s future is uncertain. The speaker wonders whether entry-level developer jobs will shrink as agents take on tasks that juniors previously handled. They also note that legacy systems and organizational inertia may sustain substantial work for years.
• The approach is not yet conclusive. The speaker mentions contrasting views from Habr commenters, some of whom reportedly argue that this approach does not work well over the long term. They consider their current experience promising, but not definitive.

Reviews, guides, or tutorials

This is a personal account and analysis of an AI-assisted development workflow, not a step-by-step tutorial or formal product review. It discusses OpenCode, OpenSpec, and GLM through the speaker’s workplace experience.

Main speaker/source: The video’s host, a software developer describing their team’s experience. Habr commenters and reports about other companies are mentioned secondhand.