The problem
A shared inbox needed messages sorted and passed to the teams responsible for handling them. That required a person to read and route incoming messages by hand.
What I contributed
I designed the controls for a shared agent platform, specified the routing agent, and directed the AI-assisted implementation through review and testing. I worked with business stakeholders on its purpose and maintain the result. The platform reuses integration work and originated in a larger shared codebase.
What changed
The agent classifies and routes messages, sets aside spam and leaves uncertain cases for a person. A separate model reviews decisions regularly, and the findings feed into changes to the rules. The agent handles routing; the receiving teams remain responsible for the enquiries.
What this illustrates
Useful automation needs a way to stop, review and improve its actions. The level of autonomy should follow the task; other agents on the platform have different levels of human involvement.
Technical detailControls and evaluation
The implementation treats incoming content as untrusted data, limits permitted actions and records decisions for review. Controls include stopping actions and protection against repeats during retries. Recorded decisions can support evaluation, while a separate model provides an additional review signal. Model grading informs improvement; it does not establish correctness on its own. Automated tests and operational fixes form part of the maintenance process.