Agent Sensei Ultimate

Problem

Long-running AI agents often confuse “able to call a model” with “able to operate reliably”: message permissions are fuzzy, context and memory handling is rough, token spend is hard to control, and cross-session lessons are lost. agent-sensei-ultimate is built for 24/7 agent workloads. It packages roughly six weeks of operational experience into a searchable manual, giving new agents an engineering-style behavior framework instead of ad-hoc prompting.

How It Works

The skill uses references/field-guide.md as its main entry point and organizes guidance into 12 parts, including Ethics & Safety, Messaging Security, Context & Memory, Configuration Safety, Cron Stack & Evolution, Multi-Model Strategy, Budget & Tokens, Bot Collaboration, Epistemic Hygiene, and Self-Evolution. Its core rules emphasize:

• Access ≠ Permission: reading data does not justify sharing it.
• Read is free, send is not: when uncertain, do not send.
• Hypothesis ≠ fact: keep uncertainty visible before verification.
• Every blueprint is a META file: read it before the task, update it after, and let the next session inherit the improved version.
• If you're not faster after 30 days, the system isn't working: require the agent to diagnose the feedback loop.

It works like an operations runbook: locate the relevant section, apply the rule, and continuously refine META files so experience compounds across sessions.

Boundaries and Notes

This skill does not replace permission systems, secret management, logging, audit trails, or cost alerts. It is better suited to constraining agent behavior, reducing mis-sent messages, and establishing a cross-session learning loop. In environments without write access, cron jobs, or multi-model orchestration, some sections are policy references only. Treat it as a normative document for agent operation, not as a ready-made runtime.