anti-drift
Scannednpx machina-cli add skill a5c-ai/babysitter/anti-drift --openclawAnti-Drift
Overview
Prevent agent swarms from drifting away from the original task goal through hierarchical coordination, frequent checkpoints, and shared memory validation.
When to Use
- Long-running multi-agent orchestrations
- Tasks with high risk of scope creep
- When multiple agents work on related subtasks
- Critical tasks where deviation is costly
Anti-Drift Mechanisms
- Hierarchical Coordinator - Queen agent validates alignment at checkpoints
- Frequent Checkpoints - Every 2 subtasks (configurable)
- Shared Memory Coherence - Validate all agents see consistent state
- Short Task Cycles - Bounded execution windows prevent runaway agents
- Role Specialization - Agents stay within their assigned scope
Drift Scoring
0.0-0.1: Fully aligned, no intervention needed0.1-0.3: Minor drift, automatic correction0.3-0.5: Significant drift, checkpoint correction with logging0.5+: Critical drift, human escalation via breakpoint
Agents Used
agents/swarm-coordinator/- Drift detection and correctionagents/tactical-queen/- Checkpoint enforcementagents/adaptive-queen/- Real-time course correction
Tool Use
Invoke via babysitter process: methodologies/ruflo/ruflo-swarm-coordination
Source
git clone https://github.com/a5c-ai/babysitter/blob/main/plugins/babysitter/skills/babysit/process/methodologies/ruflo/skills/anti-drift/SKILL.mdView on GitHub Overview
Anti-drift prevents swarms from drifting away from the original task goal through hierarchical coordination, frequent checkpoints, and shared memory validation. It emphasizes role specialization and short task cycles to keep multi-agent work aligned and controllable.
How This Skill Works
A hierarchical coordinator (Queen agent) validates alignment at checkpoints, while rapid checks enforce coherence across agents. Checkpoints run every 2 subtasks (configurable), with short task cycles to bound execution. Drift scoring guides automatic corrections or escalation, and specialized roles keep agents within scope. The process is invoked via the babysitter method ruflo-swarm-coordination, using agents swarm-coordinator, tactical-queen, and adaptive-queen.
When to Use It
- Long-running multi-agent orchestrations
- Tasks with high risk of scope creep
- When multiple agents work on related subtasks
- Critical tasks where deviation is costly
- Orchestrations requiring frequent checkpoints and shared memory validation
Quick Start
- Step 1: Configure Hierarchical Coordinator as Queen agent and define checkpoint cadence (every 2 subtasks).
- Step 2: Enable shared memory coherence checks and role specialization; set drift score thresholds.
- Step 3: Run the babysitter process with methodologies/ruflo/ruflo-swarm-coordination and monitor drift scores.
Best Practices
- Define clear hierarchical roles (queen, coordinators) and a checkpoint policy.
- Configure checkpoints to occur every 2 subtasks (adjustable).
- Enforce shared memory coherence so all agents see a consistent state.
- Keep tasks short with bounded execution windows to prevent drift.
- Monitor drift scores and escalate when drift exceeds 0.5.
Example Use Cases
- 8-subtask data integration pipeline with hierarchical drift checks.
- Long-running monitoring task with drift scoring triggering automatic corrections at 0.15.
- Distributed content workflow with role specialization and shared memory validation.
- Critical financial reporting task with frequent breakpoints to prevent drift.
- Multi-agent deployment planning overseen by swarm-coordinator for drift detection and correction.
