Lately, I’ve been experimenting with OpenCode and reading up on multi-agent development architectures. However, both my own tests and the demos I’ve seen leave me wondering about their real value.
The big promise of these architectures (using an orchestrator alongside specialized sub-agents) is minimizing or even completely eliminating human intervention. While the potential is there, the cost seems too high. In my tests, relatively simple requirements consistently resulted in massive Pull Requests costing upwards of €40 to €50 each.
A common pattern in these setups is essentially replicating a full human team virtually: sub-agents acting as Project Manager, Technical Product Owner, UI/UX designer, Architect, Developer, QA, and so on. While this allows for very smart context management, copying human team structures probably isn't optimal for AI, and the financial cost is huge.
Given this, I struggle to see a clear advantage over simpler approaches, such as using a single solid coding agent selected according to the task, or maybe two agents using an adversarial pattern (like builder and reviewer).
What has your experience been? Do you think we’re over-engineering things?
Top comments (0)