github.devotd.
NewAI agents

100yenadmin/hermes-teammates v1.0.0 — Named Subagent Teammates

A look at the features and configuration boundaries of 100yenadmin/hermes-teammates, a plugin that calls subagents by name in Hermes Agent and carries execution forward.

Repository100yenadmin/hermes-teammates ↗

Named subagent teammates for Hermes Agent: an operator-defined roster, context-carried follow-ups, live steering and a Kanban teammates lane. Stock Hermes; no patched core.

stars
★ 79
Forks
0
Language
Python
License
MIT
Latest release
v1.0.0
Read original source ↗View on GitHub ↗

100yenadmin/hermes-teammates is a Python plugin that adds named subagent teammates to Hermes Agent. Once the operator configures teammates, the model picks a name, issues instructions, and the plugin handles execution. The repository's latest release is v1.0.0, published under the MIT license. It uses the public plugin API without modifying the Hermes core. The available information does not list any changes specific to v1.0.0.

Teammate configuration covers a description, instructions, a tool list, and the optional model, route, and reasoning_effort. Follow-up instructions carry over a summary of the previous instructions, and you can steer a running teammate by passing it the sentence for the next step. It also provides a Kanban teammate column. The execution context allocates space within a 30,000-character limit in the order: operator instructions, card body, follow-up summary, and caller info, and it reports the character count of any truncated part. The card body is capped at 6,000 characters.

Configuration is checked for format, but not for whether values are actually usable in the environment. Unusable values appear under Config problems in /teammates, and unknown keys, type errors, or an empty tool list block execution. If the tool list is missing or null, the parent's tools are inherited. Descriptions over 300 characters and instructions over 4,000 characters are truncated, but execution continues. The available information has no performance measurements, so real latency or cost is hard to judge, and the v1.0.0 change list cannot be confirmed.