Agent Briefing Gateway

Agent-Briefing-Gateway is a documented design model for an orchestration agent that mediates communication between a human operator and a batch of concurrently working AI sub-agents. Rather than letting each worker interrupt the human directly, the gateway collects typed events, triages and deduplicates them, applies an interrupt budget, and composes a single scheduled brief per cycle. A separate override channel handles genuinely urgent items, with urgency requested by workers but arbitrated by the gateway. The project is explicitly a model only; no code is built, and the repository exists to specify how the system should behave and to document unresolved questions.

The design rests on three rules: workers never address the human directly, a worker that needs a decision checkpoints and parks rather than holding a live context, and an approval binds to a specific proposed action so that the gateway re-verifies preconditions before resuming and executing. The documentation covers the event taxonomy, state and checkpoint resumption through an approval ledger, channel policy including quiet hours and escalation, candidate transports such as email, messaging platforms and voice calls, and a wire-format event contract. It positions the work as human supervisory control in the sense of Sheridan and Verplank's levels of automation, specifically level 5 management by consent rather than level 6 management by exception. The author describes the project as derived from an earlier outbound email batching gateway built for client communication.

The model may be of interest to developers building multi-agent systems who need to limit supervision overhead, to researchers working on human-in-the-loop or human-on-the-loop control of autonomous processes, and to engineers evaluating how approval gates and resumable agent state should be designed. The roadmap identifies checkpoint and resume on a single approval gate as the step that will determine whether further implementation is worthwhile.