mohammed firdous

ADK Investment Research Agent on GKE

·source

A hierarchical multi-agent system that researches companies, refines investment theses through an analyst-critic loop, and produces parallel bull/bear reports, deployed on GKE Autopilot.

ADK Investment Research Agent is a multi-agent pipeline that accepts a company name or stock ticker and returns a structured investment report. It uses Google's Agent Development Kit to coordinate a team of specialised agents: researcher, analyst, critic, and parallel case writers, running on Google Kubernetes Engine Autopilot.

What it is

A production-ready multi-agent research system with:

  • Hierarchical Orchestration: A sequential root agent drives a looping research team (up to 3 refinement iterations) before handing off to report writers.
  • Analyst-Critic Loop: An analyst proposes an investment thesis; a critic challenges it; the loop runs until the thesis is validated or the iteration cap is hit.
  • Parallel Case Writing: Bull and bear scenarios are generated concurrently as independent agent branches.
  • GKE Autopilot Deployment: The full pipeline runs as a containerised FastAPI service with Cloud Logging for observability.

How It's Built

  • Agent Framework: Google Agent Development Kit (ADK) 1.27.5 with a SequentialAgent root and a LoopAgent research team.
  • LLM Backend: Gemini 2.5 Flash for all reasoning steps.
  • Research Tool: Wikipedia API via LangChain Community integration.
  • Serving Layer: FastAPI + Uvicorn, containerised for Kubernetes.
  • Infrastructure: GKE Autopilot with Cloud Logging for structured log output.
  • State Flow: PROMPT → research → INVESTMENT_THESIS passed through shared agent state.

What I Learned

  • Multi-Agent Topology: Composing sequential, looping, and parallel agent patterns within a single ADK pipeline.
  • Feedback Loops in AI Systems: Designing critic agents that drive iterative refinement without infinite loops.
  • Cloud-Native AI Deployment: Packaging an agentic workload for GKE Autopilot and wiring up structured logging.
  • State Management: Threading structured state across agent boundaries cleanly in a stateful pipeline.
cd ..