LLM supervisor agent design #36

Closed
opened 2026-08-17 19:49:08 +02:00 by SirStone · 1 comment
Owner

Parent: #29
Blocked by: #30, #31, #32, #33

Question

How should the LLM supervisor agent be structured to drive the autonomous training loop?

Key sub-questions:

  • Interface: Claude Code skill? MCP tool? A prompt template the LLM follows?
  • tmux interaction: How does the LLM launch/monitor/restart the training runner in tmux? Direct tmux commands via bash?
  • Check interval: How often does the LLM read stats and evaluate? Time-based (every 5min)? Round-based (every 100 rounds)?
  • Action space: What can the LLM do when it detects a stall? Adjust hyperparams + restart? Change opponent? Just log and continue?
  • Session management: Does it run as a long-lived agent, or get invoked periodically by a cron/scheduler?

Depends on: all other decisions (output format, tunable params, delivery mechanism, stall detection strategy).

Parent: #29 Blocked by: #30, #31, #32, #33 ## Question How should the LLM supervisor agent be structured to drive the autonomous training loop? **Key sub-questions:** - **Interface**: Claude Code skill? MCP tool? A prompt template the LLM follows? - **tmux interaction**: How does the LLM launch/monitor/restart the training runner in tmux? Direct tmux commands via bash? - **Check interval**: How often does the LLM read stats and evaluate? Time-based (every 5min)? Round-based (every 100 rounds)? - **Action space**: What can the LLM do when it detects a stall? Adjust hyperparams + restart? Change opponent? Just log and continue? - **Session management**: Does it run as a long-lived agent, or get invoked periodically by a cron/scheduler? **Depends on:** all other decisions (output format, tunable params, delivery mechanism, stall detection strategy).
SirStone added the wayfinder:grilling label 2026-08-17 19:49:08 +02:00
Author
Owner

Out of scope for this map. The training runner is a standalone CLI tool; the LLM supervisor is a separate effort that simply invokes, monitors, and restarts this tool externally.

Out of scope for this map. The training runner is a standalone CLI tool; the LLM supervisor is a separate effort that simply invokes, monitors, and restarts this tool externally.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: SirStone/SirRoboGarage#36