Case study / 04
Policy and Briefing Research Agent
One policy idea, four audiences, three models. An agentic flow that splits judgment, grunt work, and craft across Opus, Haiku, and Sonnet — and exposes every step in the UI.
Phase / 01 Discover
Map the workflow.
The Institute for Progress publishes the same policy idea four ways — for Hill staff, journalists, industry executives, and the public. Each audience needs different framing, different evidence, different register. I studied how their writers research once and rewrite many times, and landed on a clean separation: research should be audience-agnostic, audience adaptation should happen only at generation.
Phase / 02 Build
Ship the system.
Three-step agentic flow, one model per job. Opus 4.6 reads the policy and proposes 2–4 research questions (judgment work). Parallel's search API gathers excerpts and Haiku 4.5 summarizes them (grunt work, parallelized). Sonnet 4.5 writes the final brief in IFP's voice with the audience-specific style guide injected (craft work). Each step is exposed in the UI so the user can edit research questions before execution — the agent's decisions are legible, not hidden.
Phase / 03 Deploy
Earn trust.
Packaged as a Render Docker web service with a second "Build with AI" tab — a markdown editor with on-demand source surfacing — for longer-form drafting. The deployment target was a working demo IFP could open, click through, and stress-test, not a production app for end users.
Phase / 04 Compound
Reuse the learning.
The reusable insight is architectural: don't pay for one big model to do every step when three smaller models can each do what they're best at, and don't pay twice for orchestration when the workflow is already legible. The same pattern now sits behind LegiKit's bill-summarization fallback (CRS → Gemini → OpenAI) and shapes how I structure any new AI workflow — pick the cheapest model that can hold the job, and expose the seams.