ActualCode

Overview

ActualCode targets a specific hiring problem: LeetCode tests abstract algorithms, not real codebases. Candidates who ace LeetCode often struggle with actual repositories.

The system analyzes any GitHub repository and generates a realistic, implementable coding assessment in about two minutes. Input a repo URL and difficulty level, and seven specialized agents collaborate to produce a problem that matches the tech stack, code patterns, and architectural decisions of that codebase.

Built with Google Vertex AI, the Agent Development Kit (ADK), and the Agent-to-Agent (A2A) protocol. 2nd place at the Google Cloud × AI Tinkerers Hackathon. First time building with Google’s agent ecosystem.

Architecture

The system orchestrates 7 agents through A2A protocol, each with explicit contracts and responsibilities:

• Agent 1 - Scanner: Fetches repository data via GitHub API, handling rate limiting and pagination for large codebases.
• Agents 2-5 - Parallel Analysis (single-pass optimization):
• Code Analyzer (Gemini 2.5 Pro): Analyzes architecture, patterns, and tech stack
• PR Analyzer (Gemini 2.5 Flash): Reviews recent pull requests to understand current development focus
• Issue Analyzer (Gemini 2.5 Flash): Examines open issues to identify real weaknesses Dependency Analyzer (Gemini 2.5 Flash): Maps tech stack and dependencies for feasibility checking
• Agent 6 - Problem Creator (Gemini 2.5 Pro): Synthesizes analysis into a realistic coding problem with starter code, requirements, acceptance criteria, and hints.
• Agent 7 - QA Validator (Gemini 2.5 Flash): Validates problems across 4 dimensions (feasibility, quality, technical match, educational value) with automated scoring and specific improvement feedback.
• A2A Protocol Implementation: Agents communicate via structured messages with typed contracts. The orchestrator manages sequencing, parallel execution, and error recovery without hardcoding flows.

Key Learnings

1. A2A enforces clean agent boundaries. Each agent has to define explicit input/output schemas and capabilities. The constraint feels restrictive but produces maintainable, composable agents that can be reused across workflows. The protocol’s “skill” declarations make intent discoverable.
2. Repository-specific beats generic. Problems generated from real codebases land differently. When the assessment references actual patterns from the repo (React error boundaries for facebook/react, middleware architecture for expressjs/express), candidates engage more deeply.
3. Single-pass analysis needs careful orchestration. Running five agents in parallel risks overwhelming context or producing race conditions. The orchestrator has to manage token budgets, merge analysis streams, and hand the Problem Creator a coherent synthesis. Rough split for this project: 30% agent design, 70% coordination logic.
4. QA validation has to be multi-dimensional. A good coding challenge needs clear requirements, must match the repo’s stack, and has to test relevant skills. A single score misses too much. The 4-dimension rubric (feasibility, quality, technical match, educational value) catches failure modes a single metric does not.

Status

Launched September 2025 at the Google Cloud Hackathon. Currently processing live repositories via CLI.