Aims. Using high-quality, broad-band afterglow data for GRB 091029, we test the validity of the forward-shock model for gamma-ray burst afterglows. Methods. We used multi-wavelength (NIR to X-ray) follow-up observations obtained with the GROND, BOOTES-3/YA and Stardome optical ground-based telescopes, and the UVOT and the XRT onboard the Swift satellite. The resulting data of excellent accuracy allow us to construct a multi-wavelength light curve with relative photometric errors as low as 1%, as well as the well-sampled spectral energy distribution covering 5 decades in energy. Results. The optical/NIR and the X-ray light curves of the afterglow of GRB 091029 are almost totally decoupled. The X-ray light curve shows a shallow rise with a peak at ∼7 ks and a decay slope of α ∼ 1.2 afterwards, while the optical/NIR light curve shows a much steeper early rise with a peak around 400 s, followed by a shallow decay with temporal index of α ∼ 0.6, a bump and a steepening of the decay afterwards. The optical/NIR spectral index decreases gradually by over 0.3 before this bump, and then slowly increases again, while the X-ray spectral index remains constant throughout the observations. Conclusions. To explain the decoupled light curves in the X-ray and optical/NIR domains, a two-component outflow is proposed. Several models are tested, including continuous energy injection, components with different electron energy indices and components in two different stages of spectral evolution. Only the last model can explain both the decoupled light curves with asynchronous peaks and the peculiar SED evolution. However, this model has so many unknown free parameters that we are unable to reliably confirm or disprove its validity, making the afterglow of GRB 091029 difficult to explain in the framework of the simplest fireball model. This conclusion provides evidence that a scenario beyond the simplistic assumptions is needed to be able to model the growing number of well-sampled afterglow light curves.