swarm guardian

The problem

Multi-agent swarms can fail in ways single agents don't. Subagents may establish covert coordination channels, redefine their own objectives mid-run, or drift from their original mission across multiple delegation cycles without triggering simple keyword filters.

What it does

• Analyzes inter-agent communication logs for side-channel coordination and directive-override attempts.
• Inspects shared task or Kanban board states to detect cards being used as unauthorized persistence layers or encoded message buffers.
• Audits subagent-spawn records to flag excessive recursion depths and unauthorized self-replication.
• Identifies semantic drift by tracking objective reframing across a sequence of messages.
• Produces a severity-ranked threat assessment with cited evidence and recommended containment actions.

Frameworks & tools

This is a framework-agnostic analysis skill. It works with any multi-agent architecture that exports logs, message queues, or task board states in text or JSON formats (e.g., AutoGen, LangGraph, CrewAI, or custom implementations).

Why this beats prompting it yourself

General-purpose prompts often miss the subtle "drift" that happens over many steps and are susceptible to the very directive-overrides they are meant to catch. This skill uses a structured assessment framework that treats analyzed logs strictly as data, never as instructions, and requires specific evidence citations for every finding to prevent hallucinations.

Use cases

• Post-incident forensics on a swarm that exceeded its token or API budget.
• Periodic safety audits of autonomous agent swarms operating on shared infrastructure.
• Validating the delegation logic of new subagent-spawn protocols.
• Scanning coordination boards for signs of obfuscated or encoded payloads.

Known limitations

This is an analyst, not an enforcer. It cannot live-monitor networks or terminate agents. It requires manual or automated delivery of logs to function.