Deep Research Agent
LiveMay 2026 – Present
developer-tools
Built to comprehensively demonstrate the OpenAI Agents SDK feature set: handoffs, input/output guardrails, agent-as-tool, structured Pydantic outputs, streamed orchestration via Runner.run_streamed(), the built-in WebSearchTool, a custom @function_tool for email delivery, per-agent ModelSettings, and tracing via OpenAI's platform. A coordinator routes vague queries through a clarification step; a planner decomposes the topic into five parallel searches; a writer synthesises a 1000-word markdown report, which an email agent formats as HTML and delivers via Resend. Deployed as a Gradio app on Hugging Face Spaces with a custom domain.
Design Decisions
Handoff-based agent topology rather than a static chain — the coordinator picks the next agent at runtime, enabling clarification of vague queries without adding latency to clear ones. Guardrails run as a lightweight LLM classifier (input) and a deterministic word-count/section check (output) so unsafe queries and low-quality reports short-circuit the pipeline before downstream cost is spent.
Highlights
- ›Six agents wired via SDK handoffs: coordinator → planner → 5× search → writer → email, with an optional clarification step for vague queries
- ›Input safety guardrail (lightweight classifier agent) plus deterministic output guardrail (word-count and section checks) short-circuit the pipeline on bad queries or low-quality reports
- ›Agent-as-tool pattern: SearchAgent is exposed back to WriterAgent as a `search_web` tool for follow-up lookups during synthesis, demonstrating the SDK's tool composition primitive
Timeline
AI / LLM
Infrastructure
MLOps
Tech Stack
Metrics
6 agents + 2 guardrails
5 parallel web searches
Architecture
Deep Research Agent — Pipeline & Handoffs