Deep Research Workflow

Single-pass LLM research assistants tend to be expensive, brittle, and hard to trust for analytical work that spans multiple sources.

The challenge was to design a system that could vary its depth based on the prompt while still validating quality before delivery.

The workflow uses an explicit router and planner to decide whether a request needs lightweight synthesis or a deeper multi-agent research path.

Evaluation is treated as a first-class step rather than an afterthought, allowing the system to retry or refine when the answer quality is below threshold.

Async execution enables parallel search and evidence gathering, which materially reduces latency on the heavier research paths.

Pydantic schemas keep agent handoffs typed so downstream stages can reason over predictable payloads instead of parsing free-form text.

The main improvement came from matching the workflow depth to the task instead of forcing every request through the same expensive pipeline.

The evaluation loop made the system more production-like because it surfaced quality as an operational concern rather than a manual review task.

Agent systems become easier to evolve when their interfaces are explicit and typed. The orchestration layer matters as much as the prompts.

A future iteration would add richer telemetry per stage so routing and evaluation thresholds can be tuned from observed behavior instead of static heuristics.