Marie is built around an existing agent engine, Hermes. Rather than reimplementing its tools for each task, the project is connecting those tools to a common execution and progress interface.
Different work needs different execution
Some tasks need a model. Others need a precise calculation, document recognition, or transcription tool. The development work connects local inference and specialist executors so an agent can choose an available capability.
Making execution observable
Capability runs record progress, timing, partial results, cancellation, and errors. These records help distinguish time spent in a model from time spent doing the actual work.
Integration checks have covered synthetic document recognition and transcription inputs. They show that the components connect; they do not establish accuracy across arbitrary real-world documents.
The next constraint
Heavy local work shares limited compute resources. Scheduling, cancellation, and memory pressure remain part of the product’s reliability work.