network-ai_skill

Overview

This skill implements a local, sandboxed multi-agent swarm orchestrator for complex workflows. It decomposes tasks into three specialized agents, manages a shared blackboard for state, and enforces permission walls and budget checks before sensitive operations. All execution is local and audit-logged for compliance.

How this skill works

On receipt of a complex request, the orchestrator decomposes it into exactly three sub-tasks (DATA, VERIFY, RECOMMEND) and delegates each to a specialized agent using a budget-aware handoff interceptor. Every sessions_send is gated by a handoff check and permission scripts; agents write results to a local blackboard file. Before committing results, the orchestrator runs validation, supervisor review, and only writes a final approved state to the blackboard and audit log.

When to use it

• Coordinating parallel analysis, risk checks, and strategic recommendations for a single business problem
• Delegating work to specialized agent sessions (data_analyst, risk_assessor, strategy_advisor)
• Accessing sensitive systems (SAP_API, FINANCIAL_API) where permission checks and tokens are required
• Maintaining a persistent shared state among agents via a local blackboard
• Running budgeted workflows where token and cost limits must be enforced

Best practices

• Always initialize a task budget before delegation and check remaining budget before each handoff
• Run the handoff interceptor script before every sessions_send and abort if blocked
• Limit permission scopes and include clear justifications when requesting resource access
• Use the blackboard for explicit read/write coordination and clean up TTL entries regularly
• Validate agent results and run supervisor-review before committing final outputs to the blackboard
• Log every sensitive action to the audit_log.jsonl to preserve an audit trail

Example use cases

• Q4 financial analysis: extract metrics, run compliance checks, and produce board-ready recommendations
• Data export gating: request permission, anonymize PII, and write approved export to a local file
• Parallel risk assessment: run data processing, compliance verification, and strategy synthesis concurrently
• Incident investigation: coordinate a data pull, verify integrity, and recommend remediation steps