Agentic Claude - a deep dive
Question
- I am planning to automate a number of workflows in an enterprise.
- The key characteristics are:
- MCP tools
- Negotiate with the tools
- Plan the calling sequence
- Test for validation during and after work
- Should I use the "agent sdk" that has the built in loop, or should i create my own "harness and loop"?
Claudes opinion
- Use the Agent SDK's built-in loop.
- It's purpose-built for exactly what you're describing?enterprise workflows with MCP tools, planning, and validation.
Reason
- Built-in agentic loop: Claude automatically plans the calling sequence, negotiates with tools, and adapts based on results. You don't reinvent this.
- MCP integration: Connect external tools (databases, APIs, etc.) directly. The loop handles tool orchestration.
- Validation hooks: Use PreToolUse and PostToolUse hooks to intercept and validate tool calls before and after execution?exactly what you need for testing during and after work.
- Cost and turn limits: Cap spending with maxBudgetUsd and maxTurns to prevent runaway workflows.
- Permissions system: Control which tools Claude can use and require approval for sensitive operations.
when does your own harness makes sense....
- Building your own harness makes sense only if you need fundamentally different behavior (e.g., a non-agentic orchestration model, custom message types, or a completely different loop structure).
- For enterprise automation with negotiation and validation, the SDK's loop is battle-tested and handles the hard parts.