BossAgent ADK Production Error Investigation

Overview

This document captures the current forensic analysis of the four reported production failures and related architectural gaps in the BossAgent ADK deployment on Google Cloud. Evidence is pulled from the active codebase to highlight misalignments with the intended 32‑cell Cloud Run architecture, Vertex AI Claude usage, and frontend data integrity.

Error Summaries and Evidence

1) Audio Capability Failure with Gemini Fallback

• The BossAgent FastAPI wrapper and runtime use the SimpleBossAgent, which calls Gemini directly and never initializes Claude/Anthropic on Vertex AI. There is no path for passing transcribed audio text into a Claude pipeline, so the voice hand‑off cannot succeed. 【F:src/agents/adk/app.py†L13-L169】【F:src/agents/adk/boss_agent_simple.py†L25-L170】
• Only four cells (3, 26, 27, 28) are configured in the running agent, leaving the remaining 28 cells unreachable during orchestration. 【F:src/agents/adk/boss_agent_simple.py†L38-L44】

2) Complete Chat Failure with API Errors (400/500)

• Chat endpoints delegate directly to SimpleBossAgent.process_message; any exception becomes a 500, and lack of initialized agent yields a 503. There is no validation or routing to the broader ADK stack, which explains generic 400/500 patterns instead of structured cell/tool calls. 【F:src/agents/adk/app.py†L71-L131】
• Because the agent only knows four cells, commands like ping cell3 can never exercise the full 32‑cell network, and missing URLs for other cells likely result in failed downstream calls when requests reference them. 【F:src/agents/adk/boss_agent_simple.py†L38-L44】

3) MOA Agent Deployment/Launch Failure

• The current runtime is built around a lightweight Gemini caller and does not consume the richer BossAgentConfig that enumerates additional cell URLs or supports Anthropic. Invocation of pre‑deployed cells is limited to a small hard‑coded map; there is no discovery or invocation path for the remaining cells, which encourages fallback behaviors such as ad‑hoc deployment attempts. 【F:src/agents/adk/boss_agent_simple.py†L38-L44】【F:src/agents/adk/config.py†L16-L64】
• Cloud Build style deployment scripts remain under cell directories, reinforcing the outdated deploy‑on‑demand pattern rather than invoking pre‑deployed Cloud Run services. 【F:src/cells/cell20/Cell 20 - Dynamic Configuration Management†L1-L1】【F:src/cells/cell32/cell32-causal-roi-dashboard/create_deployment_package.py†L1-L5】

4) Frontend Mock Data Instead of Real Data

• Remote specialized agents fall back to mock responses whenever the backend call fails, marking responses with mock: true, so production pages can silently display fabricated data. 【F:miz-oki-command-center-ui/agents/specialized/RemoteAgentBase.ts†L18-L60】
• The MOE router explicitly uses mock embeddings (// Get query embedding (mock for now)), preventing real routing scores and reinforcing mock behavior throughout agent selection. 【F:miz-oki-command-center-ui/agents/routing/MOERouter.ts†L14-L120】

Cross-Cutting Architectural Gaps

• Vertex AI Claude integration is absent in the active FastAPI service, preventing the mandated Claude Sonnet 4.6 primary path and forcing Gemini-only fallbacks.
• There is no central registry or invocation layer for all 32 cells; only four are wired, so orchestration, health checks, and ping style commands cannot cover the full architecture.
• Frontend data paths tolerate backend failures by defaulting to mock payloads, which masks outages and leaves users seeing placeholder information instead of live cell status.

Immediate Remediation Priorities

1. Replace SimpleBossAgent with a Claude-first Vertex AI client that accepts transcribed audio text and routes it into the chat pipeline, using the correct model ID (claude-sonnet-4-6@20260201) and global endpoint.
2. Introduce a complete cell registry (all 32 Cloud Run URLs) and switch BossAgent orchestration to invoke those services instead of attempting deployments.
3. Remove or gate mock fallbacks in the frontend agents/MOE router and ensure production builds call live APIs with observable failures.
4. Align the FastAPI error handling to return actionable diagnostics (including cell target and model path) instead of generic 400/500 responses.