Software teams have long repeated that a mistake costs more the later you fix it, up to a hundred times more once a product ships. The idea pushed teams to decide as much as possible up front. After four days with an agent team, I think the cost of a mistake has moved somewhere else.
Code got cheap to change
The evidence from the project is plain. On day two, one development session merged 13 app features. In one night, the agents changed the hook that guards their lanes eight times. When I changed the shape of the whole team on day four, the same sessions took their new roles within the hour.
The old claim was already weaker than its reputation. A study of 171 software projects found no consistent rise in the cost of fixing an issue later. With agents that rewrite, retest and verify code in hours, the curve gets flatter still.
Some things stayed expensive
Cheap code did not make everything cheap. In the project, the costs showed up in other places.
| Cheap to change now | Still expensive |
|---|---|
| Rewriting a module, or a guard, many times | What users see and feel, once they have seen it |
| Changing how the team is organised | Real data, once customers trust you with it |
| Trying a design in code and throwing it away | My attention: one unanswered question held four pull requests for four hours |
| Adding tests after a review finding | Review and verification: most of 237 hours on app pull requests went to stacks, review loops and handoffs |
What users see is the hardest to undo, and for Bluecliff it matters most, because the product is about trust. If an answer drafted by AI is wrong and a customer sees it, the code can be fixed within the hour. The customer's trust in the next answer takes much longer to come back. That is why Bluecliff marks every AI draft and waits for a person to confirm it.
What I decide first now
- What users see. I designed the product, the UX and a design system before choosing the tech stack. The design is also what the agents check their work against.
- How real data is handled. The data model and who can see what deserve care before customers arrive, because migrating trusted data stays slow and risky.
- Where my attention goes. A short list of decisions that need me, and nothing else waiting on me.
Everything else, I let the agents try, change and throw away. Most of it is cheaper to rewrite than to plan in detail.
Open questions
I don't know when code becomes expensive again. It may happen with the first real customers, with a security audit, or with a codebase too large for an agent to hold in its context. I'll note it in the log when it happens.
The bottom line
With agents, the cost of a mistake moved from the code to the people who meet it: the users, the customers who trust the data, and the one person who has to check the work. I decide those parts first, and let the agents move fast on the rest.
Further reading
- Menzies and others, Are delayed issues harder to resolve?
- An org chart made of agents, for the design-first choice