Hermes builds a personal wiki and Slack digests

Razorpay engineers describe an actual employee's Hermes instance after eight weeks in their internal platform. Its 21 scheduled jobs ingest information from 23 Slack channels, six GitHub repositories and Google Workspace hourly, plus selected external feeds daily. This is a particular employee's configured workflow, rather than a claim that every employee uses the same sources.

A synthesis pipeline turns the collected material into wiki pages about people, projects, decisions, threads and concepts. An entity graph and a Titan-v2 vector index support questions across that information. The instance posts a morning digest at 07:15 and an evening digest at 19:00 to a private Slack channel, and refreshes a task tracker every two hours.

Each employee's agent runs in a separate Kubernetes namespace with encrypted storage and scoped identity. Outbound requests pass through an audit proxy. Risky commands use approval flows, and deployments can optionally require review of newly written skills. These are controls described by the platform's developers, not independently audited guarantees. The article's session totals also do not reconcile, so this entry avoids using them as a productivity measure. The demonstrated useful pattern is to turn explicitly scoped work sources into a persistent, searchable record and deliver concise reports at defined times, with platform controls around access and execution.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Blog · ashwath kumar, with Siddharth Tripathi