Running it as a workshop

A proven flow that tells the whole story in ~15 minutes:

1. Run stage 1. Toggle markers on/off. “See - without communication, no ant trails.” (stigmergy)
2. Start stage 2. Chaos at first, low collection rate. Then watch the curve climb and the policy table fill. “Nobody told it the rules.”
3. Compare policies. Show “Policy match”. “The machine found the stage-1 code itself.” → That is the transition from algorithm to AI.
4. Stage 3. Show the pulsing think-ring (latency), read the human-readable thought aloud, point at the ghost-call counter. “Reliability has to be measured.”

No participant needs to read code. Yet everyone has seen what an agent is, what learning is, and where an LLM helps - and where it doesn't.

Extension ideas

• Beetles + a fighter caste - introduces the WirdAngegriffen conflict (fight vs. flee) and role specialisation. Closer to the original AntMe!.
• A second colony - competition for the same sources. A bridge to real multi-agent systems.
• Model benchmark - run stage 3 with different local models and compare ghost-call rate, latency and collection rate. At its core already a mini-benchmark for agentic reliability.

The essence

You didn't build “an ant simulation”. You built one loop and swapped the decision box four times. That is precisely the mental map for every agent system:

The environment (world, tools, data) is decoupled from the decider (rule, network, LLM). “AI” is the swappable box inside a loop that existed before it - not the loop itself.

Once you've seen this with your own eyes, no AI hype will fool you again.

→ Back to the lab