QCon London AI Coding State of the Game: More Capable, More Expensive, More Dangerous Coding Agents

2026-03-22 · Source: InfoQ · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cybersecurity & Data Privacy · Depth: Advanced, short

Summary

Birgitta Böckeler, Distinguished Engineer at Thoughtworks, presented a QCon London keynote on March 22, 2026, detailing the evolution of AI coding from "vibe coding" to autonomous agents and swarms. She highlighted context engineering as a significant advancement, noting Anthropic's shift from monolithic rules files to granular "skills" for lazy loading, which optimizes context window usage. Böckeler observed a move towards "hands-off" coding, with agents running unsupervised for up to 20 minutes and integrating directly into CI/CD pipelines via GitHub Actions. However, she cautioned about the rising costs and a worsening security landscape, particularly due to prompt injection attacks, citing an incident where an attacker extracted secrets via a crafted GitHub issue. She also introduced a risk framework for agent supervision based on error probability, impact, and detectability.

Key takeaway

For AI Architects evaluating autonomous coding agents, recognize that while these tools offer increased automation and integration with CI/CD, they demand robust security protocols and careful cost management. Your teams must define and enforce strict practices for agent supervision, especially regarding untrusted content and external communication, to mitigate prompt injection risks and prevent the amplification of poor coding practices.

Key insights

AI coding agents are more capable but introduce significant security risks and increased operational costs.

Principles

• Context engineering improves agent performance.
• Unsupervised agents amplify existing practices.
• Security is a conceptual, not technical, problem.

Method

Supervise AI coding agents using a risk framework considering error probability, impact, and detectability, with granular context management via "skills" and lazy loading.

In practice

• Implement granular agent "skills" for context.
• Sandbox unsupervised agents rigorously.
• Integrate agents with CI/CD pipelines.

Topics

• AI Coding Agents
• Context Engineering
• Prompt Injection
• Software Security
• Agent Swarms

Code references

• wilsonzlin/fastrender

Best for: CTO, AI Architect, VP of Engineering/Data, AI Engineer, Software Engineer, AI Security Engineer

Related on AIssential

• Stop giving agents the whole computer — AI coding agents are advancing in autonomy and capability, but demand stricter security and structured development workflows.
• Three AI coding agents leaked secrets through a single prompt injection. One vendor's system card predicted it — AI agent runtime environments, not just models, are critical attack surfaces for prompt injection vulnerabilities.
• Presentation: State of Play: AI Coding Assistants — AI coding assistants are evolving towards greater autonomy, necessitating advanced context engineering and robust safety harnesses.
• How to Optimize Your AI Coding Agent Context — Effective context management is paramount for optimizing AI coding agent performance and engineer efficiency.
• Claude Code, Codex and Agentic Coding #8 — AI coding agents are rapidly advancing but require careful management to mitigate risks from autonomous actions.
• AI Coding Agents, Deconstructed — Effective AI agent systems require structured architectures that separate concerns to manage context and scale workflows.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by InfoQ.