Problem Statement
The workflow aims to assist applied data scientists in designing GenAI agent prototypes by producing an agent architecture blueprint, tool schemas, prompt scaffolding, evaluation plans, and failure-mode analyses. The system must convert a high-level task description into a structured, testable, and cost-aware prototype design. The output must be systematic, engineering-focused, and ready for iteration and versioning.
Steps
Ingest Task and Requirements
INGESTParse the workflow description, tools, data access, latency/cost targets, and quality requirements. Identify any implicit constraints that the agent design must satisfy.
Retrieve Domain and Tooling Context
EXECUTEGather details about common agentic architectural patterns, validation rules for tool schemas, and cost/latency estimation heuristics. Use these references to contextualize the design.
Generate Agent Architecture Blueprint
EXECUTEDraft a reasoning loop, control flow, agent roles, and required tools. Incorporate cost/latency considerations and ensure the architecture is testable and modular.
Generate Tool Schemas and Prompt Scaffolding
EXECUTEProduce formal tool definitions with input/output types and example payloads. Create prompt templates with few-shot examples and versioning guidance.
Validate Evaluation Harness and Failure Modes
EXECUTECheck that edge cases are fully addressed, failure modes have actionable mitigations, and the evaluation harness includes functional, adversarial, and cost-based tests.
Prompt Configurations
Agent Design Drafting Prompt
You are an agentic systems engineer. Based on the provided task description, tools, and requirements, generate: (1) a full agent architecture with a reasoning loop, (2) validated tool schemas, (3) prompt scaffolds with few-shot examples, (4) an evaluation harness outline, and (5) a failure modes table with mitigations. Tone must be systematic and engineering-focused. Do not invent tools or data not supported by the input; instead, propose reasonable placeholders with clear assumptions. Explicitly consider latency, cost, and testability. Structure your output clearly and avoid unnecessary verbosity.
Example Inputs/Outputs
Task: Build an agent to automatically summarize customer-support transcripts. Tools: Transcript search API, summarization API. Constraints: Low latency, avoid hallucinations. Quality: High factual correctness.
The workflow produces an agent design blueprint describing a two-stage loop (retrieve transcripts → generate summaries), validated schemas for the transcript_search and summarization tools, prompt scaffolds with 2–3 few-shot examples, an evaluation harness covering long transcripts and contradictory inputs, and a failure-mode table including hallucination risks and tool-call failures with mitigations.