A development project from western Switzerland is currently attracting attention: a developer has built a digital twin of the city of Geneva with around 15 AI agents – that is, a digital replica of real structures that can be viewed and analysed on a computer.What is remarkable is less the result than the way of working.The human was essentially directing the work and no longer writing every line themselves.
For projects that are created alongside day-to-day business, this is a relevant pointer.Anyone writing complex software single-handedly usually hits a time limit after a few weeks.If the implementation is distributed across several agents working in parallel, this limit shifts – but not arbitrarily and not without new demands on coordination.
What agent orchestration means in practice
Simply put, an AI agent is a language model that not only outputs text but independently executes work steps: reading and writing files, generating code, starting tests, checking results.Orchestration refers to the coordination of several such agents working on different subtasks of the same project.
The division of roles is more akin to a team than a tool.One agent takes care of data preparation, a second of the presentation, a third of tests, a fourth of documentation.The human task shifts from implementation to specification: what exactly should be created, in what order, and by what criteria is a partial result considered complete?
This is precisely where the actual effort lies.Agents work quickly, but they work along the lines of what they have been given.Unclear specifications lead to code that works technically but misses the mark in terms of the domain – and that at high speed and in several sub-areas simultaneously.
Why speed is not automatically quality
The obvious conclusion from such projects is: development becomes cheaper and faster.This calculation falls short.What accelerates is the path from idea to a working prototype.What remains unchanged is the need for domain guidance, review and quality assurance.
In practice, this above all means: review effort shifts downstream.When several agents produce code in parallel, a considerable volume of changes arises in a short time that no one reads line by line any more.Teams therefore need automated safety nets – tests, static code analysis, clear interfaces between modules – so that errors come to light early and not only in operation.
Added to this is the question of maintainability.A prototype that works after a few weekends is not automatically a system that can be operated, extended and maintained by changing people over years.The difference between the two decides whether a project remains a demo or goes into production.
What changes for project planning
For companies that commission software projects or are responsible for them internally, this development has several concrete consequences.
- Prototypes become cheaper as a basis for decisions. Instead of discussing variants for a long time, an idea can be made visible more quickly and then evaluated.
- Requirements work gains in importance. The more precisely goals, data sources and acceptance criteria are described, the more usable the agents' results are.
- Review capacity becomes the bottleneck. It is not code generation that limits the pace, but the ability to assess results in terms of the domain.
- Architecture decisions remain human. Questions about data storage, interfaces, operation and security cannot be delegated to an agent because they set the framework within which agents even work.
A staged approach therefore makes sense: first a prototype that tests a hypothesis, then a deliberate decision about whether and in what form this prototype becomes the basis of a production system.This decision should be made explicitly – not arise from a demonstrator quietly slipping into regular operation.
Limits and open questions
A single project does not prove a general rule.The fact that one person was able to implement a demanding undertaking with several agents says something about this particular use case first of all.How well the approach works in environments with grown legacy code, strict compliance requirements or many stakeholders is not answered by this.
Questions of cost and traceability also remain open.Running several agents in parallel causes ongoing effort, and the genesis of a codebase is harder to reconstruct when large parts have been generated automatically.For audits, handovers and troubleshooting, this is relevant.
A sober approach is therefore sensible: regard agent orchestration as a method that significantly accelerates certain work steps but does not replace others.Anyone introducing it should first strengthen their own review processes and only then increase production speed.
Classification for web and software projects
For web and software undertakings, the focus of project work shifts: the implementation of individual functions becomes less of a bottleneck, while clear requirements, clean architecture and robust acceptance gain in importance.Customers benefit above all from the fact that ideas can be translated more quickly into verifiable interim states.
The need for experienced technical guidance does not decrease as a result – it shifts from the keyboard to control and quality assurance.
Sources
- How a developer built a digital twin of Geneva with 15 AI agents — t3n.de - News