Skip to main content

Claude Agents

A Claude-based security analysis agent system with tool use and multi-turn reasoning across 3 agents. The orchestrator analyzes tasks and selects tools via Anthropic, the runner executes code_scanner and dependency_checker tools then generates fix suggestions, and the evaluator assesses fix quality and coverage.

Environment variables

This example requires ANTHROPIC_API_KEY, WAXELL_API_KEY, and WAXELL_API_URL. Use --dry-run to run without any API keys.

Architecture​

Key Code​

Tool selection via Anthropic and security scanning tools​

The orchestrator uses Anthropic to decide which tools to run, then the runner executes them.

@waxell.decision(name="select_tools", options=["code_scanner", "dependency_checker", "both"])
async def select_tools(query: str, anthropic_client) -> dict:
    response = await anthropic_client.messages.create(
        model="claude-sonnet-4-5-20250929", max_tokens=512,
        messages=[{"role": "user", "content": f"Decide which tools to use. Request: {query}"}],
    )
    return json.loads(response.content[0].text)

@waxell.tool(tool_type="function")
def code_scanner(path: str, scan_type: str = "security") -> dict:
    """Scan source code for security vulnerabilities."""
    return MOCK_SCAN_RESULTS

@waxell.tool(tool_type="function")
def dependency_checker(manifest_path: str = "requirements.txt") -> dict:
    """Check project dependencies for known CVEs."""
    return MOCK_DEPENDENCY_RESULTS

Evaluator with @reasoning fix assessment​

The evaluator assesses whether all critical issues are addressed in the suggested fixes.

@waxell.reasoning_dec(step="fix_assessment")
async def assess_fixes(fixes_text: str, scan_results: dict) -> dict:
    issues = scan_results.get("issues", [])
    critical_count = sum(1 for i in issues if i["severity"] == "critical")
    high_count = sum(1 for i in issues if i["severity"] == "high")
    return {
        "thought": f"Evaluating fixes for {len(issues)} issues "
                   f"({critical_count} critical, {high_count} high).",
        "evidence": [f"{i['type']} in {i['file']}:{i['line']}" for i in issues],
        "conclusion": "All critical issues addressed" if critical_count <= 1
                      else "Multiple critical issues need review",
    }

waxell.score("fix_quality", 0.88, comment="coverage of critical issues")
waxell.score("coverage", 0.92, data_type="float", comment="issues with suggested fixes")

What this demonstrates​

  • @waxell.observe -- parent-child agent hierarchy with automatic lineage
  • @waxell.step_dec -- task analysis recorded as execution step
  • @waxell.tool -- code scanner and dependency checker tools
  • @waxell.decision -- tool selection via Anthropic Claude
  • @waxell.reasoning_dec -- fix quality assessment chain-of-thought
  • waxell.score() -- fix quality and coverage scores
  • waxell.tag() / waxell.metadata() -- provider (anthropic), tools executed metadata
  • Auto-instrumented LLM calls -- two Anthropic claude-sonnet calls captured automatically
  • Claude Agents pattern -- security analysis with code scanning and fix generation

Run it​

# Dry-run (no API keys needed)
cd dev/waxell-dev
python -m app.demos.claude_agents_agent --dry-run

# Live (real Anthropic)
export ANTHROPIC_API_KEY="sk-ant-..."
python -m app.demos.claude_agents_agent

# Custom query
python -m app.demos.claude_agents_agent --query "Review authentication flow for vulnerabilities"

Source​

dev/waxell-dev/app/demos/claude_agents_agent.py