Switchboard / Hermes
A Telegram operations agent that gates real actions with allowlists, durable claims and explicit verification.
Problem
An agent that can post messages or search the web needs strict rules around who may trigger an action, where it may run and whether a retried update has already been handled.
How I solved it
- 01
Checks exact Telegram users and channels before allowing an action, so authority is enforced at the entry point rather than inferred later in the agent flow.
- 02
Claims each inbound update durably before handling it, preventing retry delivery from producing duplicate external work.
- 03
Continued the buildathon slice with Convex traces, a sanitised dashboard, tests and GitHub Actions verification so action decisions remain inspectable.
My role / what I owned
Buildathon builder and repository maintainer
- Implemented exact user and channel allowlists around Telegram actions.
- Added durable claiming so an inbound update is handled once across retries.
- Continued the repository after the initial buildathon slice with traces, tests and verification automation.
How it was built
- Team
- A Hermes Buildathon entry; no separate collaborator breakdown is documented in the public repository.
- Tools
- TypeScript
- Telegram
- Convex
- GitHub Actions
- Use of AI
- Hermes agent tooling is part of the system. There is no documented percentage split for AI coding assistance.
Source and current state
- 53 public commits at the 9 August 2026 review.
- 8 actual test modules and 1 GitHub Actions verification workflow.
- Convex traces and a sanitised dashboard are included in the public repository.
- Status
- The initial buildathon slice was built in an afternoon; the current public repository reflects substantial later work.
- What it shows
- The public repository exposes the action gates, durable claims, traces, test modules and CI workflow.