QCon London: AI for developers is in a 'dangerous state'

"AI is in a dangerous state where it is too useful not to use, but where by using it, developers are giving up the experience they need to review what it does. Context engineering involves curating the information that models or agents see to get better results, including rules, commands, instructions, and resources like MCP tools that enable more accurate task performance."

"Even though context windows are a lot bigger now than a year ago, when they get full, the effectiveness of the agent degrades and it starts costing a lot more money. Sub-agents allow a main agent to spawn other agents to perform specialized tasks and report back, reducing load on the main agent and enabling separate context windows for code reviews or different models."

AI development tools present a paradox: they are indispensable yet risky for developer growth. Context engineering has emerged as a key focus, involving careful curation of information, rules, and resources that models receive to improve accuracy. Model Context Protocol (MCP) tools enable local definition of resources, reducing remote context size and costs. As context windows expand, managing their fullness remains critical to maintain agent effectiveness and control expenses. Sub-agents represent an advancing feature where main agents delegate specialized tasks to independent agents with separate context windows, reducing load and enabling code reviews with different models. The industry trend moves toward reduced supervision, with developers potentially building custom agent orchestrators.

Read at Theregister