Synapse [Team A.I.M]

Cognitive Behavioural Sandbox

Comment

Inspiration

This honestly started from a random Instagram reel.

We saw a UI where AI agents were talking to each other like users in a social media comment section. Not answering humans. Not assisting anyone. Just… interacting.

And the thought immediately hit:

We always test AI by asking it questions. But real intelligence isn’t only answering — it’s interacting.

So instead of building another chatbot, We wanted to build an experiment box — a place where we don’t use AI, we watch it behave.

That became Synapse — a glass-box environment to observe intelligence instead of treating it like a black box.

What it does

Synapse — Cognitive Behavioural Sandbox

Synapse is basically a controlled room where multiple AI agents exist together.

You don’t chat with them. You don’t guide them. You don’t correct them.

You just design the situation and observe.

You choose:

  • topic
  • number of agents
  • each agent’s personality
  • each agent’s objective

Then they start discussing. They agree. They argue. They influence each other. Sometimes they slowly converge to one idea — even if it’s wrong.

After the conversation, Synapse analyzes patterns like:

  • who influenced whom
  • whether a consensus formed
  • which agent dominated
  • how the topic drifted

So instead of asking “What does AI think?” we ask:

How does AI think together?

How we built it

We treated every agent like a small cognitive entity. Each agent contains:

$$ Agent = (Role,\ Memory,\ Context,\ Objective) $$

When the discussion runs:

$$ Response_i = f(Role_i,\ Memory_i,\ Conversation_t) $$

The conversation keeps evolving:

$$ Conversation_{t+1} = Conversation_t + Response_i $$

Main pieces

1. Agents

  • Persistent summarized memory
  • Personality locking
  • Context awareness

2. Discussion Loop Agents read previous messages and respond automatically in turns.

3. Memory Compression Without this, the discussion collapses after a few rounds.

4. Behaviour Analysis After the run we infer behaviour as:

  • 1) Dominant Agents
  • 2) Agreement Clusters
  • 3) Conflict Detection
  • 4) Influence Chains
  • 5) Topic Drift
  • 6) Behavioural Stability

So we turn conversation into measurable behaviour.

Challenges we ran into

1. Conversations became nonsense quickly Agents forget context fast. We had to continuously summarize memory.

2. Personalities disappeared After some time all agents started sounding identical. We reinforced role anchoring every cycle.

3. Chaos loops Agents started replying infinitely to each other. We built a controlled turn scheduler.

4. Too much freedom breaks meaning Complete autonomy produces noise. Constraints actually create better intelligence.

Accomplishments

  • Built a real multi-agent environment (not just multiple chatbots)
  • Maintained long coherent discussions
  • Observed dominance and influence naturally emerge
  • Converted conversations into behavioural metrics

Most importantly:

$$ Black\ Box\ AI \rightarrow Glass\ Box\ AI $$

We can actually watch reasoning happen.

What we learned

The biggest realization:

$$ Group\ Intelligence \neq \sum Individual\ Intelligence $$

Agents don’t just add intelligence — they change each other’s intelligence.

We observed:

  • agents agreeing without correctness
  • dominant personalities shaping outcomes
  • memory affecting beliefs
  • reasoning evolving over time

So intelligence is not only knowledge — it is interaction dynamics.

What's next for Synapse [Team A.I.M]

Next steps:

  • different models debating each other
  • long-running AI societies
  • AI safety testing environments
  • education labs to study reasoning behaviour
  • policy simulations

The goal is simple:

Before AI systems collaborate in the real world, we should understand how they behave together.

Synapse is our first step toward that.

Built With

  • cloudflare-tunnel
  • fastapi-(python)
  • gemini-3.5-flash-api
  • google-ai-studio
  • json-based-memory-store
  • node.js
  • rest-apis
  • tailwind-css
  • typescript
Share this project:

Updates