Georgetown Wicked Problems Lab: a classroom where AI assists but never decides

The situation

Georgetown's School of Foreign Service teaches security studies students to reason about wicked problems. These are complex policy challenges where the right answer depends on judgment, framing, and willingness to sit with ambiguity. Off-the-shelf AI tools are eager to give students the answer. That's exactly the wrong behavior for teaching policy reasoning. An AI that writes the student's analysis teaches the student nothing except how to paste.

The faculty wanted something different. They wanted a teaming platform where the AI helps with the mechanical parts of analysis (retrieving sources, surfacing counterexamples, flagging weak evidence) while leaving every real judgment where it belongs.

What we built

The Wicked Problems Lab. Student teams work through a structured four-phase sequence. The phases are problem articulation, refinement, solution planning, and pressure testing. Each team has a project, a chat surface, and an orchestrator running in the background. The orchestrator listens to the team's conversation (text, and optionally voice via a separate WebSocket server that routes audio to Gemini Live), classifies intent, and decides which of eight specialized analysis agents to dispatch.

The agents run in three tiers. Fast-tier agents (Gemini 2.5 Flash) handle intake, triage, framing. Deep-tier agents (Gemini 2.5 Pro) do synthesis and analysis. QA/QC-tier agents (Gemini 2.5 Pro at low temperature) challenge the team's reasoning, surface gaps, and run adversarial review. Every agent returns structured output validated by a Zod schema, gets scored for importance, and surfaces to the team only if the score exceeds a threshold. Instructors get cross-team read access, voice controls, and intervention tools.

Behind the whole thing: Supabase with row-level security scoping students to their team, PostgreSQL stored procedures enforcing phase transitions, pgvector embeddings on voice transcripts, and realtime subscriptions driving the UI.

How it's defensible

The platform's most important code is the smallest. It's a delegation contract, a set of rules written in TypeScript that constrains what the AI is allowed to do inside the learning loop.

The AI cannot verify the team's assumptions. If a student states an assumption, the system records it as an open assumption and surfaces it for pressure testing. The AI never marks an assumption as "verified" on the student's behalf.

The AI cannot define success criteria. Students decide what a good answer looks like for their project. The AI can suggest criteria to consider, but it cannot write them into the project record.

The AI cannot advance the team past a phase gate. Phase gates are PL/pgSQL stored procedures. The platform itself enforces the constraint. An agent cannot call "you're done, move on." Only the team's human decisions, captured through specific UI actions, can trigger a transition.

Those three rules sound simple. They change the whole shape of the product. A classroom AI that respects them becomes a teaching tool. One that doesn't becomes an answer-generator, and the class might as well be a ChatGPT subscription.

What it replaced

Two alternatives, both bad. The first is keeping AI out of the classroom entirely, which misses the chance to teach students how to work with these tools in a professional setting. The second is letting a generic assistant write the analysis, which misses the point of the class. The Wicked Problems Lab is the third option: AI in the room, under discipline.

What a similar engagement looks like

Fourteen to eighteen weeks for a first deployment. We need faculty time to define the phase sequence, the success-gate criteria, and the agent library for the domain. You get a deployed platform with the orchestrator, the agent library, the gate enforcement, instructor tooling, and a voice subsystem if the classroom wants it.

It fits universities, executive education programs, professional training, and any setting where the point is to develop human judgment, not substitute for it. It's the wrong product if you want an AI that does the work and hands you the answer.