Methodic

Autonomous research operations agent that turns B2B business decisions into governed, evidence-linked data.

Track 1: Build — Net-New Agents | Google for Startups AI Agent Challenge

What Methodic Does

Static B2B surveys produce shallow answers like "price." Methodic replaces them with an autonomous multi-agent workflow:

1. Plans the study — organizer + methodology agents structure objectives and push back on sampling bias
2. Conducts adaptive interviews — probes vague answers into specific decision variables in real time
3. Triangulates against CRM context — MCP tools pull deal and telemetry data mid-interview
4. Tracks coverage and autonomously re-plans — identifies gaps across 8 canonical research variables
5. Exports structured data — evidence-linked rows with confidence scores directly to BigQuery

Result: +0.692 composite quality improvement over static surveys (same rubric, same participants; coverage from 16.7% to 100% across 8 research variables).

Live Demo

Cloud Run: https://methodic-2030382823.us-central1.run.app

• Demo UI: /static/demo.html — watch the full pipeline run autonomously
• Interactive mode: /static/interactive.html — participate as the interview subject
• Agent card: /.well-known/agent-card.json

Quick Start (Local)

pip install -r requirements.txt
uvicorn methodic.server:app --port 8080
# Open http://localhost:8080/static/demo.html

Architecture

SequentialAgent (root)
├── plan_phase (SequentialAgent)
│   ├── session_init (BaseAgent)
│   ├── organizer (LlmAgent)
│   ├── methodology (LlmAgent)
│   └── question_design (LlmAgent)
├── fieldwork (LoopAgent, max_iterations=3)
│   └── fieldwork_body (SequentialAgent)
│       ├── interviewer (LlmAgent)
│       ├── participant_sim (LlmAgent, demo mode)
│       ├── extractor_step (BaseAgent)
│       ├── turn_checker (BaseAgent)
│       ├── coverage_assessment (BaseAgent)
│       └── replan_step (BaseAgent)
└── finalize (SequentialAgent)
    ├── quality_reviewer (LlmAgent)
    ├── bigquery_export (BaseAgent)
    └── completion (LlmAgent)

Stack: Google ADK · Gemini 3.1 Pro Preview · MCP (stdio JSON-RPC 2.0) · FastAPI · Cloud Run · BigQuery · Vertex AI

Project Structure

methodic/                  # ADK agent package
  __init__.py              # Model configuration (Gemini 3.1 Pro / Flash)
  agent.py                 # Agent graph (SequentialAgent + LoopAgent)
  server.py                # FastAPI server (SSE streaming, interactive API)
  schemas.py               # Pydantic models (ParticipantResponse, 8 canonical fields)
  agents/                  # Custom BaseAgent steps (extraction, coverage, re-plan, BQ export)
  tools/                   # MCP client, BigQuery export, quality scoring
  static/                  # Demo UI (demo.html) + Interactive UI (interactive.html)
scripts/                   # MCP server, validators, recording, legacy WP scripts
fixtures/                  # Deterministic test data (CRM, telemetry, participants)
tests/                     # 73 unit/integration + 60 Playwright E2E tests
docs/                      # Submission materials, evidence, reviews

Tests

# All tests
python3 -m pytest tests/ -v

# Unit/integration only
python3 -m pytest tests/ --ignore=tests/e2e -v

# E2E (requires running server)
python3 -m pytest tests/e2e/ -v

Evidence

Evidence	Status	Source
Cloud Run health	PASS	curl /health
133 tests passing	PASS	pytest --collect-only
Live SSE stream (25–34 events)	PASS	live-run-2026-05-14.md
BigQuery live export (3 rows)	PASS	bq query on methodic_demo.win_loss_responses
Agent card	PASS	/.well-known/agent-card.json
Quality delta (+0.692)	PASS	Fixture benchmark

Honest Labels

See docs/limitations.md for what the prototype proves and what it does not.

Submission Materials

• Devpost Copy
• Demo Script
• Limitations
• Architecture Diagram