StudyHour

StudyHour Watch - autonomous GCSE exam agent that monitors the open web, verifies sources, maps updates to revision tasks, publishes a cited briefing, $ access w x402, & logs action 2 ClickHouse.

Comment

Inspiration

GCSE students and teachers do not just need another chatbot. They need to know when official assessment rules change, which sources to trust, and what to practise next.

StudyHour Watch was inspired by a real pain point: exam-board and regulator updates are scattered across Ofqual, AQA, OCR, Pearson Edexcel, WJEC and other pages. A policy change like formulae and equation sheets affects how teachers plan revision and how students practise exam questions. We wanted an autonomous agent that could watch the open web, verify official sources, and turn an assessment update into an actionable StudyHour revision plan.

What it does

StudyHour Watch is an autonomous GCSE assessment-intelligence agent.

It searches the open web for GCSE policy updates, verifies official sources, extracts teacher and student actions, maps the change to StudyHour’s revision corpus, publishes a cited brief to Cited.md, exposes a payment-ready full brief, and logs every agent decision into ClickHouse.

For the demo query, it tracks GCSE 2026 formulae and equation sheet policy across AQA, Edexcel, OCR and WJEC. It then creates a practice plan covering equation selection, rearranging formulae, units and substitutions, and timed calculation practice.

How we built it

We built a multi-agent pipeline inside StudyHour’s Hono/TypeScript API.

The agents are:

  • Scout searches the open web using Tavily and official fallback sources.
  • Verifier accepts only trusted GCSE regulator or exam-board sources.
  • Policy extracts the decision, affected subjects, teacher actions, student actions and misconceptions.
  • Mapper connects the policy to StudyHour notes and real past-paper questions.
  • Publisher writes cited.md and publishes through Senso/Cited.md.
  • Commerce emits an x402-style payment requirement for the full machine-readable brief.

The frontend is a Vite/React demo cockpit with two tabs: Open Web Watch and ClickHouse. The Watch tab streams the agent run live. The ClickHouse tab shows every event row, including agent names, tool calls, latency, source ids, publish status and payment-gate events.

Challenges we ran into

The hardest part was making the system feel like a real agent doing real work, not just search plus a summary.

We had to make source verification explicit, show agent boundaries clearly, and ensure the UI exposed the actual orchestration. Senso publishing also introduced real-world failure modes: sometimes a new item is created as a pending draft rather than immediately returning a live Cited.md URL. We handled that by showing the draft status, Senso content id, and a latest-live Cited.md fallback while still logging the failed publish event in ClickHouse.

Another challenge was balancing demo reliability with real infrastructure. The system touches live web search, local StudyHour search, Senso publishing, payment-gate metadata and ClickHouse logging, so observability became essential.

Accomplishments That We’re Proud Of

We built a real end-to-end autonomous workflow.

The agent does not stop at answering a question. It searches, verifies, maps, publishes, monetizes and logs. It produces a public cited artifact and a queryable audit trail.

We are especially proud of the StudyHour mapping step. The agent turns an official policy update into a concrete revision sequence using real StudyHour notes and past-paper questions. That makes the output useful for both teachers and students.

We are also proud that ClickHouse is not decorative. Every event in the timeline is a row that can be queried, analyzed and debugged.

What We Learned

We learned that multi-agent demos only land when the responsibility of each agent is visible.

“Many agents” is not enough. Each agent needs a reason to exist: Scout finds, Verifier trusts, Policy interprets, Mapper connects, Publisher ships, Commerce gates, and ClickHouse audits.

We also learned that publishing is different from generation. Once an agent’s output leaves the chat window and becomes a public artifact, reliability, citations, fallback states and audit logs matter much more.

What’s Next For StudyHour

Next, StudyHour Watch can monitor more GCSE assessment domains: specification updates, Ofqual qualification conditions, exam-board policy changes, grade-boundary updates and subject-specific support materials.

We want to turn this into a continuous assessment-intelligence layer for schools, tutors and revision platforms. A teacher should be able to ask, “What changed for my Year 11 class, and what should they practise this week?” and get a verified, cited, StudyHour-backed answer.

We also want to complete the x402 payment flow so other agents can pay for full machine-readable briefs, and expand ClickHouse analytics to show which policy changes generate the most student confusion and practice demand.

Share this project:

Updates