A 15-skill multi-agent analysis engine
An R&D system I built to pressure-test multi-agent orchestration on a genuinely hard problem: investment analysis, where the inputs are noisy, the dimensions compete, and a single model's verdict is rarely trustworthy on its own. Fifteen skills, five subagents, one orchestrator.
The engagement
Ask one model to "analyze this" and you get confident mush. The real problem decomposes into independent dimensions — technical, fundamental, sentiment, risk, valuation — each wanting its own method and its own data. Run them in sequence and it is slow; run them with no scoring discipline and you have five opinions and no decision.
The orchestrator routes each request to the relevant skills, runs the heavy analysis agents in parallel, and synthesizes their output into a composite score on one 0–100 scale with documented weights. Every skill is dual-mode — it works standalone or inside an orchestrated run — so adding a dimension means adding a skill, not rewiring the system.
The framework runs fifteen skills today and is built to take more. It is also a working reference for the orchestration patterns I bring to client AI work: parallel execution, composite scoring, and skills that compose — instead of a monolith that has to be rewritten every time the question changes.
Practice areas applied
- AI Orchestration & Implementation. Design and deploy intelligent multi-agent systems that coordinate AI services, optimize workflows, and accelerate development — from Claude and Azure OpenAI integration to custom knowledge bases and automated pipelines.