← सभी गाइड पर लौटें

A Practical GPT and Claude Tool-Calling Workflow

Build One Tool Workflow Before Adding an Agent

Function calling lets GPT or Claude decide when your application should perform a real action, such as looking up an order, querying a database, or creating a support ticket. The model does not directly access your systems. Instead, it returns a structured tool request, your server validates and executes it, and then your server sends the result back to the model for a useful final response.

A practical starting workflow is an order-support assistant with two tools: get_order_status and create_support_ticket. This is safer and easier to debug than starting with ten tools or a fully autonomous agent. Once the loop works reliably, you can add inventory, billing, or CRM integrations.

Step 1: Define Narrow, Testable Functions

Design each tool around one clear business operation. For example, get_order_status should accept an order_id and return shipping state, estimated delivery date, and tracking URL. Do not create a vague tool named manage_orders with a free-form instruction field. Narrow tools reduce ambiguity and make authorization rules much easier to enforce.

• Use descriptive names: get_order_status is clearer than order_lookup.
• Document constraints: state that order_id must be a customer-visible ID, not an email address or database query.
• Keep outputs small: return the few fields the model needs, rather than an entire internal record.
• Separate read and write actions: ticket creation, refunds, and account changes should never share a generic tool.

Step 2: Create a Strict JSON Schema

Both GPT and Claude rely on a schema that describes a tool's inputs. Require fields whenever possible, define types precisely, and use enums for finite choices. A create_support_ticket tool might require order_id, issue_type, and customer_summary, while issue_type is restricted to delivery, damaged_item, missing_item, or billing.

In OpenAI-style function calling, tools are commonly sent as function definitions with a JSON Schema parameters object. The model returns tool_calls containing a function name and JSON arguments. In Claude's native Messages format, a tool definition uses an input_schema, and Claude returns a tool_use content block. The concepts are the same, but the response objects and follow-up message format differ.

Step 3: Route Requests Through 59API

Your application can use 59API as the model gateway while keeping an OpenAI SDK integration. Set the SDK base URL to https://api.59api.com, provide your 59API key from a server-side environment variable, select the supported GPT or Claude model you need, and send your normal chat completion request with the tools array attached. This approach is especially useful for teams that want a consistent integration surface while choosing the best model for each task.

59API provides pay-as-you-go access to native official-quality GPT and Claude models, including Claude Opus, Sonnet, Haiku, and Fable options, without a model downgrade. For tool-heavy applications, where requests often require multiple model turns, its low-cost relay pricing can make prototyping and production testing substantially more practical. It also offers a referral rebate for teams that share the service with other developers.

Step 4: Implement the Tool Execution Loop

After sending the user message and tool definitions, inspect the assistant response. If there is no tool request, display the normal text answer. If the model requests get_order_status, parse the arguments, validate them against your own schema, confirm the authenticated user may access that order, and call your internal order service.

Next, send the result back in the conversation with the matching tool-call identifier. In OpenAI-style flows, this is a tool message tied to tool_call_id. In native Claude flows, return a tool_result block tied to tool_use_id. Then call the model again. The model can translate raw system data such as shipped, carrier UPS, and eta 2025-04-18 into a customer-friendly answer.

Never let model-generated arguments bypass server checks. The schema helps, but it is not an authorization system. Treat every tool argument as untrusted input, enforce tenant boundaries, validate IDs, apply rate limits, and require explicit confirmation before irreversible actions such as refunds or cancellations.

Step 5: Handle Multiple Calls and Failures

A model may request several tools in one turn, such as checking an order and creating a ticket when delivery is delayed. Execute independent read-only calls in parallel if your infrastructure supports it, but preserve each call ID and return a separate result for every request. For write actions, serialize execution and add idempotency keys so a retry cannot create duplicate tickets.

When a tool fails, return a concise structured error, such as ORDER_NOT_FOUND or CRM_TIMEOUT, rather than exposing stack traces. Ask the model to explain the next safe step. It may request a corrected order ID, offer a human handoff, or retry only when your policy permits it.

Step 6: Test GPT and Claude With the Same Scenarios

Create a small evaluation set: valid order lookup, unauthorized order lookup, malformed ID, delayed shipment, unavailable backend, and a request to create a ticket. Run the same scenarios against your selected GPT and Claude models. Compare tool selection accuracy, argument quality, unnecessary calls, latency, and final-answer clarity. Log the tool name, sanitized arguments, duration, result code, and model request ID for every turn.

Keep your application logic model-neutral and place provider-specific message formatting behind a small adapter. That makes it easy to use a fast, economical model for routine lookups and a stronger model for complex support cases. When you are ready to test this workflow with affordable model access, sign up for 59API and connect your existing SDK to its API base URL.