Directing AI agents without knowing how to code: what has changed

AI has been able to write code for a while. What is new: a non-technical person can now direct it.

Until recently, having software built without knowing how to code meant outsourcing: you explained, someone else carried it out, and you discovered the result at the end. Generative AI initially reproduced this pattern, only faster. You asked, it produced, and you received a block of code that was often unreadable to anyone who did not program. Useful, but impossible to direct: you could not correct the course along the way without understanding what you were reading.

AI writes better code, certainly, but the recent change lies elsewhere: we have learned to divide its work into roles, stages and approval points, directing it like a team rather than consulting it like an oracle.

From one giant to a troupe

A single model entrusted with an entire product loses focus: it forgets a constraint given ten minutes earlier, mixes levels of detail, and heads in a direction nobody requested. The answer: share the work among several agents with narrow remits, rather than entrusting everything to one. One scopes the need, another writes the specifications. A third lays the foundations, the next builds, and the last checks.

Each agent then receives a clear role and a precise deliverable, and reads fewer methodological instructions, leaving more attention for your project itself. For you, the effect is concrete: instead of a monologue you cannot follow, you see who does what and know which stage you are at.

Putting human approval back at the centre

The other shift concerns the space left for human decisions. An agent that runs from start to finish on its own is impressive in a demonstration and unmanageable in real life: when it makes an early mistake, that error spreads everywhere. Explicit approval at every stage stops this spread. Work pauses, you read a proposal in everyday language, you correct or approve it, and nothing continues without your agreement.

That is what makes direction possible for a non-technical person. You do not need to read code to judge whether a brief reflects your intention, whether a user journey makes sense, or whether a page says the right thing. You judge the result in your own language, at a point when correcting it costs almost nothing.

Directing is not executing

One honest question remains: are you really directing if you do not understand the underlying technology? Yes, on one condition: the tool must translate its decisions into your language and leave you the final say. A conductor plays no instrument during the concert; they know the work, hear when something sounds wrong, and decide when the note stops. That is the position these tools make possible today. The skill that matters now: knowing what you want, and recognising it when you see it.