When Google announced 25% of its new code came from AI, many thought programming was dying. By 2026 the number exceeds 50-75% at global tech companies - yet developer headcount did not fall; it transformed.
When Google announced that more than 25% of the company’s new code came from AI, many took it as the end of the programming profession. By 2026 that number has surged past 50-75% at global tech companies - and something unexpected happened: developer headcount did not fall; it transformed.
What agentic development actually means
Agentic software development is not just "Copilot autocompletes your code" (that is level 1). It is an AI agent owning the entire pipeline - requirement → design → code → tests → deploy (levels 4-5). The developer shifts from "code writer" to "agent director".
What AI is good at, and what it is not
From real usage inside Devsign’s engineering teams:
- Very good: standard CRUD, unit-test generation, language conversion (e.g. Python → Go), refactoring legacy code, documentation, fixing bugs with clear error messages
- Workable: designing new APIs, complex business logic, debugging cross-system issues
- Still weak: architecting large systems, organization-wide trade-off decisions, deep business context, managing stakeholders who disagree
The real situation in Thailand
Many Thai startups already use AI to help write code. Teams of 3-4 now deliver what used to need 8-10 people. What rises in importance is code review - AI writes extremely fast, but fast is not always correct. Teams without solid review processes accumulate hidden technical debt until one day the system breaks.
The new role: from coder to architect
The developers who remain valuable over the next five years are those who:
- Understand the business deeply - translating spoken needs into systems AI can build
- Read code better than they write - reviewing ten times more AI-written code than before
- Design systems - choosing technology, setting architecture and building guardrails so AI works correctly
- Wield AI tools skillfully - knowing how to ask AI to get exactly what is needed
A real-world caution
A frequent find: AI-written code that "looks right" but hides security vulnerabilities - SQL injection, or APIs opened without authentication. Devsign’s engineering teams pair AI code generation with strict quality gates, delivering 60% faster without sacrificing quality.
