AI Agents Inside SaaS Products
The change has to be visible
A recurring bug in in-product assistants is that the tool works and nothing appears to happen. The value changes in state, but the screen does not update, because the repaint lives inside the event handler of the manual control the user normally clicks, and the programmatic path never goes through it.
The lesson generalises. If a user can do something two ways, the side effects belong in a shared function rather than in whichever handler was written first. Verify by checking what is actually painted, not by checking the model updated.
Scope it to what the product does well
An assistant that can do a few things reliably beats one that can attempt anything. Every tool you expose is a surface that can fail confusingly, so it is worth being selective and covering the paths users repeat rather than the ones that demo well.
Frequently Asked Questions
Why does my AI tool change a value but nothing appears on screen?
The repaint is usually inside the manual control's event handler, which the programmatic path skips. Move shared side effects into a common function.
How many tools should an in-product assistant have?
Fewer than you think. Each tool is a surface that can fail confusingly, so cover the paths users repeat rather than the ones that demo well.
Give your agent real tools
AetherWave runs an MCP server exposing 16 creative tools across image, video, audio and mastering, callable from any compatible AI client without a bespoke integration.
Explore the MCP server