Pathfinder

Inspiration

Many students are interested in research but do not know how to move from a broad interest to a manageable first project. Guidance is fragmented, and one-shot AI advice can sound polished while overlooking prerequisites, workload, or uncertainty.

We designed Pathfinder around a student success case conference: specialized perspectives contribute to the plan, a critic challenges unrealistic assumptions, and a synthesizer turns that discussion into an actionable pathway the student can review with a mentor.

What it does

A student provides a goal, background, weekly time budget, and real-life constraints. Pathfinder then runs a visible multi-agent workflow:

1. Goal Mapper defines what success means for this student.
2. Path Scout sequences skills, milestones, and a manageable research direction.
3. Feasibility & Safety Critic identifies overload, missing prerequisites, unsupported assumptions, and details that require human confirmation.
4. Pathway Editor reconciles those contributions into one mentor-ready Research Launch Plan.

The output includes an immediate first move, a realistic four-week sequence, opportunities to explore, questions for a human advisor, and an editable faculty outreach draft marked review before sending.

How we built it

Pathfinder uses a browser-based student intake and results experience connected to a zero-dependency Node 22 server. POST /api/plan sends the shared student profile to three role-specific GPT-5.6 Terra agents in parallel, then gives their structured reports to a fourth GPT-5.6 Terra synthesis call. The UI distinguishes live model output from a clearly labeled deterministic demo mode and presents one reconciled plan instead of a stack of chat responses.

Codex accelerated the whole development loop: product scaffolding, structured agent contracts, frontend and backend implementation, debugging, end-to-end testing, and submission documentation. The human builder made the key decisions about the student audience, agent responsibilities, case-conference metaphor, and safety boundaries.

Challenges we ran into

The hardest problem was making multi-agent collaboration materially useful rather than decorative. Each specialist therefore has a separate accountability, and the critic's findings are passed into synthesis so they can narrow or revise the final recommendation.

We also had to balance ambition with student safety. Pathfinder uses synthetic demo data, requests minimal non-sensitive context, and flags institutional details for verification instead of inventing facts.

Accomplishments that we're proud of

• A complete journey from an uncertain student goal to an actionable plan.
• Three parallel specialist agents with structured, inspectable handoffs.
• A critique step that can reduce an unrealistic scope.
• A coherent product experience that exposes collaboration without overwhelming the student.
• A clear human-review and privacy boundary for education.

What we learned

Multi-agent systems are most useful when collaboration is designed around accountable perspectives, not simply the number of agents. Product constraints also improve model behavior: a weekly time budget, a first deliverable, and explicit uncertainty produce a more useful plan than open-ended advice.

What's next

• Source-grounded retrieval from official catalogs, faculty pages, and campus resources.
• Side-by-side pathway comparison with tradeoff explanations.
• Advisor review comments and student-controlled revisions.
• Evaluation with students and educators for workload realism, usefulness, accessibility, and uncertainty calibration.
• Multilingual intake while keeping recommendations grounded in verified sources.

Safety and responsible use

Pathfinder is an educational planning aid, not a replacement for faculty or advisor judgment. It uses synthetic student information in the demo, does not require grades or student records, and directs users to verify institutional details with qualified people and official sources.

Interactive demo: https://jackal092927.github.io/pathfinder-agents/

Source code: https://github.com/jackal092927/pathfinder-agents

Built With

• codex
• css
• gpt-5.6-terra
• html
• javascript
• node.js
• openai-api