Frankstahl
Market: Manufacturing | Vertical: Steel A German back-office agent that automatically processes highly unstructured RFQs sent by customers via email and converts them end-to-end into fully priced quotes.Language: German | Channel: Email | Platform: Wonderful
View Agent
Open in Wonderful Platform
Overview
| Field | Value |
|---|---|
| Market | Steel Sellers |
| Vertical | RFQ Automation |
| Model | Gemini 2.5 Flash, multi–sub-agent architecture |
| Language | German |
| Channel |
Skills
- Frankstahl – End-to-end automation of RFQ intake, interpretation, and quote generation from free-form customer emails.
Key Tools
| Tool | Purpose |
|---|---|
input_structuring() | Parses raw RFQ emails and outputs a structured JSON describing the requested items. |
cart_configuration() | Extracts and structures customer details (name, address, email, phone, etc.). |
product_type_finder() | Classifies each requested item into one of ~350 predefined product types. |
query_by_form_and_dimensions() | Queries the items database for products matching the selected type and dimensions; returns only relevant attributes. |
variant_selection() | Selects the most appropriate item variant from the query results and returns its item ID. |
get_config_option() | Retrieves all available configuration options for the selected item ID. |
configure_variant() | Chooses the correct configuration based on the original RFQ description. |
post_quote() | Submits the finalized quote to the client’s CRM, including required metadata. |
Prompting Techniques
Decomposition into small, specialized sub-agents, where each sub-agent is responsible for a single, well-defined task. Deterministic tools are preferred wherever possible to ensure reliability and consistency. For example, the querying tools (query_by_form_and_dimensions()) are implemented as deterministic backend steps rather than LLM-driven logic, ensuring consistent and reliable product lookups.
Lessons Learned
What worked:- Back-office automation benefits significantly from decomposing complex workflows into small, focused components.
- Deterministic logic should be used whenever feasible, with LLMs handling ambiguity and interpretation.
- When the process requires a long reasoning step for multiple different items, it’s possible and recommended to run them in parallel rather than in a one-by-one queue, significantly improving throughput.
- Designing effective agentic processes required a deep understanding of the existing human workflow.
- Building a “traditional” toolkit and connecting the pipeline in the backend demands some R&D resources (±2 weeks junior dev).
- When breaking the large process into small tasks, we must also think about how to test and evaluate each part separately. This is difficult since some tasks are reliant on earlier tasks’ responses, creating dependencies that complicate isolated testing.