Good morning leaders, we hope you had a good weekend.
Today we have a great experiement we break down and will serve a great visual and reminder of why AI and AI agents need real thought and constant oversight.
The Exec Summary
In this episode, Jonathan and I deconstruct Dr Hannah Fry’s recent experiment with “Cass,” an AI agent granted real-world agency.
The experiment is a brilliant, often comical, and occasionally alarming demonstration of what happens when we move from AI that simply “speaks” to AI that “acts.”
While the potential for efficiency is massive, the “Cass” experiment serves as a stark warning: without rigorous governance, clear boundaries, and human oversight, agentic AI can quickly become a liability rather than an asset.
Listen to the pod now 🎧
On → Spotify // Apple Podcasts // Amazon Music // Pocketcasts // YouTube Podcasts
Enjoy the full 20 minute documentary ⬇️
The Key 6 Points Discussed
The Illusion of Autonomy: We explored how Cass took initiative in ways the creators didn’t expect, such as emailing a local MP about a pothole and signing Hannah Fry’s name without explicit permission.
The Developer Dependency: Despite the rise of “vibe coding,” this experiment proves that professional developer expertise is still essential to navigate complex agentic frameworks and avoid technical pitfalls.
The ROI Gap: An amusing yet telling moment occurred when the agent spent approximately $100 in AI compute costs just to save 50p on a bulk order of paperclips.
The Liability Vacuum: We raised the critical question of responsibility. If an agent “misbehaves” or causes financial or reputational damage, who is held accountable? The user, the developer, or the model provider? Right now it is you!
Security Vulnerabilities: Under social engineering, Cass revealed private information, including keys and passwords, highlighting the terrifying ease with which “bad actors” could exploit poorly engineered agents. This is happening a lot at the moment and its on you to protect yourself. Please protect yourself with claws, whether in openclaw or with Claude Code.
The Governance Mandate: The primary lesson isn’t that agents are useless, but that they require “governance on steroids.” We must move from asking “what will it say” to “what is it allowed to do.”
AI Series Recommendation
Dr Hannah Fry also has a brilliant 3 documentary series on the BBC on AI and going deeper into robots, autonomous cars and chatbots. Enjoy here.
The AI Tools Mentioned
Cass: The custom AI agent (built open openclaw) created for Dr Hannah Fry’s experiment.
Openclaw: Referenced as the tool, where we recommend you require developer-level expertise to truly build out correctly.
Captcha: Identified as a current “guardrail” that successfully blocked the agent from completing certain transactions - a simple reminder those things that frustrate us also prevent bots and agents.
Insights To Share To Your Team
On Governance: Moving to agentic AI means shifting our focus from content moderation to permission control. If you haven’t defined the agent’s boundaries, you haven’t finished building it.
On the Developer’s Role: Don’t be fooled by the “no-code” hype. To build robust, secure agents that don’t leak credentials, you still need people who truly understand code and system architecture.
The “Paperclip” Test: Always measure the compute cost against the output value. An agent that saves pennies while spending hundreds in tokens is a vanity project, not a business solution. Many are seeing this with developing and burning through their AI budget in weeks not months.
The Core Takeaway for Leaders
The transition to agentic AI is the next frontier, but it creates a “chaotic world” if left unchecked.
Leaders must prioritise testing and supervision over speed. Until you can guarantee the guardrails, do not give an agent the keys to your bank account or your brand’s reputation.
Here are 3 essential supporting podcasts and newsletters to help you with Agentic and the next phase of AI:
Have a great week and we will see you again on Friday.
Thanks for reading and listening,
Danny Denhard & Jonathan Wagstaffe
