In March 2016, something extraordinary happened in the world of strategy. AlphaGo, an artificial intelligence system created by DeepMind, faced off against Lee Sedol—one of the greatest players in the 3,000-year-old game of Go. No machine had ever beaten a top human player under professional match conditions. Most experts didn’t think it was possible.
Then came game two. On move 37, AlphaGo played a stone so unconventional, so outside traditional human play, that commentators were stunned. It wasn’t a move a human would have made—not because it was wrong, but because it was too creative. What looked like an error was, in fact, a masterstroke.
That move changed the tide of the match and ultimately helped AlphaGo win four games to one. Lee Sedol, visibly shaken but gracious in defeat, called it “beautiful” and said it had “touched the divine.”
Move 37 became symbolic—not just of AI’s power, but its potential to think differently. It showed us that AI wasn’t limited to replicating human strategies. It could invent its own. And in that lies the real opportunity: not just to be faster or cheaper, but to find smarter paths we’d never considered.
Lessons for Building AI Agents
In the conversations I’ve had as we are building AI agents, there is often fierce debate about how specific we should make the instructions to the agent.
Should you spell out the steps it needs to take to replicate the human function?
Or should you give it a clear objective and guard rails and let it find its own way there?
The safest option when you are deploying an agent is the first one, but potentially you might miss a huge opportunity.
The 2016 Alpha Go was trained by studying 100,000+ human games and then reinforcement learning as it played around 30 million games against itself. Our approach to AI Agent development will be to learn from our human best practice to start. We then allow it assess those results by always testing multiple variants of AI agents with different instructions and comparing their performance at the same point.
The unanswered question is whether we should always choose the content of it’s next test, or whether we should allow it to create it’s own set of instructions. The advantage it will have, is that it will have read every conversation that occurred on one particular trigger and will be much better placed to understand patterns in the data. It may suggest alternative segmentation of the data so that we are sending different agents to different cohorts of customers.
Every business will have their own unique AI Agents
The Greenhouse OS prospecting agent, Ethan is the core testing ground for this approach to building. As we release this to our customers, an intruiging possibility emerges. Every agent who uses this has their own account and their own version of each agent. As AI chooses new tests, with the support of the business owners, they will end up with radically different AI agents. Each business will end up with different AI’s that improve themselves, but respond uniquely to the individual agents market. This is due to the responses to the conversations being the learning ground for the change.
These AI agents, will be available exclusively to Greenhouse OS customers
If you don’t know where to start with this and would like to have a conversation about it, drop me a Whatsapp and we can talk.
If you would like to have these articles, sent straight to your inbox, just fill in your e-mail below.
Leave a Reply