OpenClaw EP4 Hackathon Results, Every Claw Needs a Shell
Security tools, piano teachers, and genome analysts - here's the results we know about from Europe's largest AI agent hackathon.
Last week we covered the UK AI Agent Hackathon EP4 as it kicked off at Imperial College London.
With 1,200 registered participants and $13,000 in prizes - students had seven days to build production-ready AI agents with OpenClaw.
Results (so far!?)
There have been to sponsor results, so not all winners have been announced (note: the hackathon ran multiple sponsor tracks) but two clear participant themes have emerged so far: security and human-AI symbiosis!
Claw for Human × Human for Claw
The organizers framed this track with a line that evokes Isaac Asimov:
“Some claws help humans. Some humans guide the claws. The best builds? They evolve together.”
Human for Claw
Carapace won the “Human for Claw” category. It protects AI agents from prompt injections and adversarial inputs - defensive armor for agents operating in production alongside humans.
“Think of it as Armour for your Claw”
As agents move from demos to deployments, the attack surface expands - Carapace addresses the question nobody wants to ask until it’s too late: how do you armor an agent against hostile input?
Claw for Human
ClawDebussy won the “Claw for Human” category. With an AI piano practice companion. The name nods to Claude Debussy (the composer, not the model - awkward). It bridges theory and performance, listening and responding as you play.
It’s time for agents to help you learn Clair de Lune.
Anyway Track
The Anyway sponsor track drew 16 qualifying submissions.
- ContentPilot (1st) — Content automation. Took home a Mac mini.
- KrumpPhysio (2nd - shared) — Healthcare/physiotherapy assistant.
- OpenCTO (2nd - shared) — Technical leadership agent.
Finalists included UPath, Continuum Discovery, ML-Trace DataPilot, and Ai Enviro — spanning bio-engineering, data and education.
Our Creative Pick
ClawBio: Your genome, locally queried
One participant made their personal genome publicly queryable through a Telegram bot built on OpenClaw and real 23andMe data:
- Pharmacogenomic summaries across 12 genes and 51 drugs
- Polygenic risk scores for 6 common diseases
- Nutrigenomics dietary recommendations
- Photo-based medication safety checks
ClawBio “18 skills, all running locally. No genetic data leaves the machine.”
This is the local-first philosophy that makes OpenClaw compelling for personal agents.
The Expert Speakers
The opening conference featured Peter Steinberger (OpenClaw creator, now at OpenAI) alongside Thomas Wolf (Hugging Face) and Emad Mostaque (former Stability AI).
Peter Steinberger on his mission:
“I’m a builder at heart… What I want is to change the world, not build a large company.”
His next goal: “An agent that even my mum can use.”
Emad Mostaque on what agents will become:
“AI agents will outperform humans on most digital tasks by 2026. But AGI won’t be a godmind. It’ll feel like your smartest coworker on Zoom or WhatsApp. We don’t need polymaths. Just competent workers getting the job done. I want cooks, not chefs.”
Thomas Wolf on building agents now:
“To build applications today it’s very interesting to focus on your verticals, maybe focus on the set of tools that you need for your agent… very general agents might come at some point.”
The consistent thread: focus on what agents can do reliably today, not what they might do eventually.
What’s next
The EP4 results shows patterns worth noting:
Security concerns. Carapace won because prompt injection defense is the next hurdle to agent adoption.
Human-AI framing matters. The “Claw for Human × Human for Claw” structure forced teams to think about direction of assistance. Is the agent serving the human, or is the human enabling the agent?
The main hackathon tracks are still rolling out results. The OpenAI surprise prize hasn’t been announced yet. We’ll update as more winners emerge.
And for builders watching from the sidelines: the questions raised here — security, human oversight, local execution, multi-agent coordination — are the ones worth solving now. Before your agent goes from demo to deployment.