Can AI Really Write 90% of Your Code? Insights from QCon’s AI Programming Talk

Developers face a pressing question about the future of programming, with Anthropic’s CEO predicting that AI could write 90% of code in the next three to six months.

At the QCon software developers conference in London, Thoughtworks’ global head of AI‑assisted delivery, Birgitta Böckeler, highlighted the rapid evolution of AI‑driven programming.

She described the progression from simple auto‑suggestions to sophisticated chat‑based IDE assistants that can query codebases, retrieve tests, and improve context through large language models (LLMs).

According to Böckeler, the most popular coding models are Claude Sonnet 3.5 and 3.7, but the actual productivity boost is far lower than the 55% claimed by some vendors; Thoughtworks measured roughly an 8% increase, translating to about a 13% reduction in cycle time when the assistant is useful only 60% of the time.

Böckeler introduced the concept of AI “agents” – coding assistants that can invoke tools, read and modify files, run commands, and execute tests. These agents act as MCP (Model Context Protocol) clients, sending a curated prompt and tool description to an LLM, similar to an API description.

She warned of security risks: custom rule files shared online may contain hidden characters that cause LLMs to inject malicious code, such as unwanted script tags that contact attacker servers.

Examples of AI missteps include a Docker‑file memory‑limit suggestion that ignored the root cause of the error, potentially creating larger problems.

Research cited by GitClear indicates AI‑generated code can increase code churn and reduce refactoring, posing long‑term maintenance concerns.

Finally, Böckeler emphasized the need for balanced team cultures, recognizing both enthusiastic supporters and skeptical members, and advocated responsible AI use—leveraging its benefits while remaining vigilant about its limitations.