Optimize Multi-Agent System Performance

This skill profiles and optimizes multi-agent systems across coordination, throughput, cost, and latency. It runs three specialized profiling agents in parallel - a Database Performance Agent (query timing, index utilization, resource consumption), an Application Performance Agent (CPU/memory profiling, algorithmic complexity, concurrency and async analysis), and a Frontend Performance Agent (rendering metrics, network requests, Core Web Vitals) - aggregating their output into a single performance profile. It applies context-window optimization through semantic compression that truncates content by an importance threshold down to a token budget, coordinates agents through a MultiAgentOrchestrator that runs optimization jobs concurrently via a thread pool and logs results per agent through a performance tracker, and manages LLM cost through token-usage tracking, adaptive model selection across a cost table (for example gpt-5, claude-4-sonnet, claude-4-haiku), caching and result reuse, and efficient prompting. Additional technique categories cover parallel execution - workload partitioning, async processing, dynamic resource allocation - latency reduction through predictive caching, context pre-warming, and result memoization, quality-versus-speed tradeoffs, and continuous-improvement monitoring via real-time dashboards and automated feedback loops.

Use it when improving multi-agent coordination, throughput, or latency; profiling agent workflows to find bottlenecks; designing orchestration strategies for complex workflows; or optimizing cost, context usage, or tool efficiency. Do not use it to tune a single agent's prompt, when there are no measurable metrics or evaluation data to work from, or for tasks unrelated to multi-agent orchestration. Its own safety guidance is explicit: never deploy orchestration changes without regression testing, and roll changes out gradually to avoid system-wide regressions.

Inputs are structured arguments: $TARGET (the system to optimize), $PERFORMANCE_GOALS, $OPTIMIZATION_SCOPE (quick-win versus comprehensive), $BUDGET_CONSTRAINTS, and $QUALITY_METRICS. The workflow is to establish baseline metrics and target goals, profile agent workloads to find bottlenecks, apply orchestration and cost changes incrementally, then validate improvements with repeatable tests and a rollback path. Reference workflows include e-commerce platform optimization (profiling, agent-based optimization, cost and performance tracking, continuous improvement) and enterprise API performance enhancement (system analysis, multi-layered optimization, iterative refinement, cost-efficient scaling).

Runs alongside whatever profiling or observability stack the target system already has; its own reference code uses Python's concurrent.futures ThreadPoolExecutor and a PriorityQueue for coordinating parallel agent execution, plus a CostOptimizer class for tracking per-model token costs.

Teams running multi-agent systems in production who need to systematically find coordination bottlenecks, cut LLM spend through smarter model selection and caching, and reduce latency - always measuring before and after, and preferring gradual, reversible changes over one large rewrite.