🤖 Engineering Deep Dive

CrewAI Multi-Agent Architecture

A 3-agent pipeline that takes a plain-English task description and ships code — Google Gemini 3.0 Pro plans it, Claude Opus 4.6 implements it via an OAuth proxy, and OpenAI o3 reviews it. Two Python files, zero YAML configs.

📋

PM Agent

Gemini 3.0 Pro

→

💻

Dev Agent

Claude Opus 4.6

→

🔍

Reviewer Agent

OpenAI o3

The Agent Team

Three agents defined in dev_team.py. Each uses a different LLM, chosen for what it's actually good at — not marketing benchmarks.

📋

PM Agent

Google Gemini 3.0 Pro

Architects the solution. Reads the project structure, analyzes existing source files, and produces a file-level implementation plan with sub-tasks, dependencies, and acceptance criteria. Uses gemini/gemini-3-pro-preview via Google AI Studio. Has full tool access (shell, file read/write, grep) to explore the codebase before planning.

💻

Dev Agent

Claude Opus 4.6

Runs via our local OAuth proxy (claude_proxy.py on port 5111) to access full Claude Code CLI capabilities. Has zero CrewAI tools — the proxy runs an internal agentic loop where Claude calls shell_command, write_file, edit_file, and others directly. CrewAI only sees the final text summary.

🔍

Reviewer Agent

OpenAI o3

Performs deep code review using openai/o3 with full tool access. Reads the PM's plan as its checklist, then independently verifies the implementation — reading files, running builds, grepping for regressions. Returns APPROVE or REQUEST CHANGES with specific line-level feedback.

Why We Built an OAuth Proxy

The hard part wasn't CrewAI orchestration — it was getting Claude Opus 4.6 to work as a CrewAI agent at all.

Problem 1: No API Key for Claude Code

Claude Code (the CLI) uses OAuth authentication stored in macOS Keychain, not standard Anthropic API keys. If you have a Claude Pro/Max subscription, your access token lives in Keychain under the service name Claude Code-credentials. There's no way to pass this to CrewAI's standard LiteLLM integration — it expects an ANTHROPIC_API_KEY environment variable.

Problem 2: CrewAI's Parser Fragility

When Claude uses tools through CrewAI's native tool system, CrewAI has to parse the tool calls and results from the LLM response. This parser frequently breaks on complex outputs — HTML files with special characters, large file contents, nested JSON. The errors are cryptic (OutputParserException) and unrecoverable.

Solution: Local Agentic Proxy

claude_proxy.py is a ~300-line Python HTTP server that presents an OpenAI-compatible /v1/chat/completions endpoint on localhost:5111. Internally, it reads the OAuth token from Keychain, calls the Anthropic API directly with tool definitions, and runs a full agentic loop — Claude calls tools, the proxy executes them locally, feeds results back to Claude, repeat for up to 100 iterations. CrewAI only receives the final text summary. No parser issues. No token leaks.

claude_proxy.py — Keychain Token Extraction def get_token(): """Read Claude OAuth token from macOS Keychain.""" with _token_lock: now = time.time() if _token_cache["token"] and _token_cache["expires"] > now: return _token_cache["token"] result = subprocess.run( ["security", "find-generic-password", "-s", "Claude Code-credentials", "-w"], capture_output=True, text=True ) if result.returncode != 0: raise RuntimeError("Cannot read Keychain") data = json.loads(result.stdout.strip()) token = data.get("claudeAiOauth", {}).get("accessToken", "") _token_cache["token"] = token _token_cache["expires"] = now + 300 # Cache 5 min return token

claude_proxy.py — Agentic Loop (simplified) def run_agent_loop(messages, model, system=None, max_tokens=16384): """Claude calls tools → proxy executes locally → feed results back.""" for iteration in range(1, 101): # Up to 100 iterations resp = call_anthropic(messages, model, system, max_tokens) tool_uses = [b for b in resp["content"] if b["type"] == "tool_use"] # If no tool calls → Claude is done. Return final text. if not tool_uses or resp.get("stop_reason") != "tool_use": text = "\n".join(b.get("text", "") for b in resp["content"] if b["type"] == "text") return text, total_input, total_output # Execute each tool call locally messages.append({"role": "assistant", "content": resp["content"]}) tool_results = [] for tu in tool_uses: result = execute_tool(tu["name"], tu["input"]) # Runs on local machine tool_results.append({ "type": "tool_result", "tool_use_id": tu["id"], "content": result }) messages.append({"role": "user", "content": tool_results}) # Loop continues — Claude sees results, may call more tools

Agent Definitions in Code

No YAML configs — agents are defined directly in dev_team.py using the CrewAI Python API.

dev_team.py — Agent Setup # 🧠 PM — Google Gemini 3.0 Pro pm = Agent( role="Project Manager & Architect", goal="Break down tasks into clear implementation plans with specific file paths and acceptance criteria", backstory="Senior tech lead with 15 years experience...", llm="gemini/gemini-3-pro-preview", tools=ALL_TOOLS, # shell, read, write, list, search allow_delegation=False, verbose=True, ) # 💻 DEV — Claude Opus 4.6 (via agentic proxy) claude_llm = LLM( model="openai/claude-opus-4-6", base_url="http://localhost:5111/v1", api_key="not-needed", # Auth via Keychain, not API key ) dev = Agent( role="Senior Full-Stack Developer", goal="Write clean, well-tested code following best practices", backstory="Expert Flutter/Dart and web developer...", llm=claude_llm, tools=[], # No CrewAI tools — proxy handles everything verbose=True, ) # 🔍 REVIEWER — OpenAI o3 reviewer = Agent( role="Senior Code Reviewer & QA Engineer", goal="Review code for bugs, security issues, performance", backstory="Meticulous code reviewer who catches issues others miss...", llm="openai/o3", tools=ALL_TOOLS, verbose=True, )

Key detail: The Dev Agent has tools=[]. This is intentional. Claude Opus 4.6 doesn't use CrewAI's tool layer at all — it calls tools inside the proxy's agentic loop. CrewAI sends it the task, the proxy runs Claude with real filesystem tools for up to 100 iterations, and CrewAI gets back a plain text summary. This completely sidesteps CrewAI's parser.

Prerequisites

What you actually need on your machine to run this.

🐍 Runtime

Python 3.13+
crewai[tools] package
pydantic, urllib3

🖥️ Platform

macOS (Keychain required for proxy)
Claude Code CLI installed & logged in
Stable internet for API calls

🔑 API Keys & Auth

GOOGLE_API_KEY (Gemini 3.0 Pro)
Claude Code OAuth (macOS Keychain)
OPENAI_API_KEY (OpenAI o3)

Running the Team

Two terminals. One command each.

Terminal 1 — Start the Proxy # Start the Claude OAuth proxy (must be running before dev_team.py) cd ~/clawd/crewai-team CREWAI_PROJECT=~/Desktop/examcertwebsite python3 claude_proxy.py # Output: # [proxy] ✅ Token: eyJhbGciOiJSUzI1... # [proxy] 📂 Project: /Users/you/Desktop/examcertwebsite # [proxy] 🔧 Tools: shell_command, read_file, write_file, edit_file, list_directory, search_files # [proxy] 🧠 Extended thinking + 100 iterations # [proxy] 🚀 Claude Code Proxy on http://127.0.0.1:5111

Terminal 2 — Run the Crew # Run the 3-agent team with a task cd ~/clawd/crewai-team CREWAI_PROJECT=~/Desktop/examcertwebsite \ python3 dev_team.py "Fix the login page layout and add form validation" # Output: 🚀 Starting AI Dev Team... 📋 Task: Fix the login page layout and add form validation 📂 Project: ~/Desktop/examcertwebsite 🧠 PM: Gemini 3.0 Pro 💻 Dev: Claude Opus 4.6 (agentic proxy) 🔍 Reviewer: OpenAI o3 ============================================================ # PM explores the codebase, creates an implementation plan # Dev receives the plan, proxy runs Claude with filesystem tools # Reviewer verifies the changes, returns APPROVE or REQUEST CHANGES

The CREWAI_PROJECT environment variable sets the working directory for all tool calls. Every file path is resolved relative to this root. Logs are written to ~/clawd/crewai-team/logs/ in real-time — you can tail -f them to watch the agentic loop progress.

Tool Architecture

Tools exist in two layers: CrewAI-managed (for PM & Reviewer) and proxy-internal (for Dev).

🔧 shell_command

Executes arbitrary bash commands in the project root. 300s timeout. Used for node build.js, git commit, npm install, etc. Output truncated to 20K chars.