My journey to co-founding entourage

Back in 2017, AI was still quite a niche area. It was more focused on deep learning, particularly Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). If you are part of that cohort, you’ll fondly remember OpenAI’s website having all of these agents chasing each other in mazes playing various games as they explored reinforcement learning.

At the time I was running miDrive, a startup that I had raised ~$10m for, focused on the EdTech sector - bringing learning to drive into the digital age (and all of the transactions that go along with it: first car, insurance, etc.). After much reflection, I had grown the company to a stage that I could feel proud about, and completed much of the innovative product work. And I couldn’t help but keep being drawn to artificial intelligence at the weekend. Consuming pretty much every YouTube video on the topic, and going way in over my head on deeply mathmatical concepts pretending to have at least grasped a bit of it.

Always being a founder, and thinking a few horizons out - I knew that AI would be the transformational technology of my lifetime. And I wanted to be living and breathing it.

That’s when I made a calculated bet to go to a large corporate (with deep pockets) and no external pressures other than focusing on learning and building my expertise amongst some of the world’s leading scientists whom we would hire to form the team. I joined a large asset manager as global head of AI products, and was tasked with infusing AI, ML, and NLP throughout the portfolio management. For the next five or six years I continued to level up my knowledge both in theory, but more importantly in practice. There were very few people building AI products at the scale we were – evident from the constant hiring battles we’d be in with the likes of Google Deepmind, Palantir, and Citadel.

Armed with deep technical knowledge and battle-tested experience deploying AI at scale, I was ready to return to my founder roots. The years as a senior executive at a large publicly traded company at the forefront of AI adoption had given me something invaluable: credibility that would matter when fundraising, coupled with insights into the real challenges enterprises face when implementing AI systems.

That journey led to entourage. I’ve always admired systems that get smarter through collaboration. Think of open-source projects where one coder’s fix helps thousands. Now apply that to AI agents. Most agents today work in silos. They tackle problems, but their lessons vanish after the task. No shared history means repeating errors, rebuilding workflows, and stalling on dynamic challenges. I see this as the big gap in AI’s promise, and why I co-founded entourage.

We started entourage to fix that. Our core idea is a shared memory protocol for AI agents. Agents capture experiences during tasks—successes, failures, patterns—and assimilate them into reusable knowledge. This isn’t offline training; it’s real-time learning embedded in the work. Retrieve what’s needed based on context, and the whole network improves. It’s like giving agents a collective hippocampus, where individual discoveries fuel group progress.

Take Spark, our first application for developers. Modern coding involves AI tools like Cursor or Copilot, but they often hit walls with platform-specific quirks. Documentation lags, and debugging feels isolated. Spark changes that. An agent in your IDE queries the shared memory for proven solutions—code patterns, workarounds, API evolutions. If it works, that success gets captured and ranked by relevance, freshness, and fit. Next time, another developer benefits. Platforms get analytics on usage, too. It’s not just assistance; it’s a network effect where every interaction makes the system better.

import requests
from api_client import APIClient

# Having issues with new API endpoint
def fetch_user_data(user_id):
    client = APIClient()
        response = client.get(f"/v2/users/{user_id}")  ❌ 404 Error
    return response.json()

import requests
from api_client import APIClient

# Fixed with Spark's help!
def fetch_user_data(user_id):
    client = APIClient()
    headers = {'X-API-Version': '2'}
        response = client.get(f"/api/v2/users/{user_id}", headers=headers)  ✓ 200 OK
    return response.json()

Then there’s Omni, which shifts agents from rigid scripts to adaptive exploration. I used to assume agents needed predefined tools to function reliably. But in unpredictable settings, that limits them—they can’t adapt to new challenges. Omni changes the game by letting agents discover solutions through code execution, direct observation, and lessons from the network. Each trial becomes shared knowledge: successes map out better paths, failures highlight pitfalls to skip. It’s multiplayer learning, where one agent’s breakthrough helps the next, turning isolated efforts into a distributed protocol of collective progress.

This opens up practical uses. For knowledge retrieval, agents draw from scattered public and private sources, building on what others have verified. In trading, they construct signal pipelines that evolve with market data, pulling from network-wide patterns to spot opportunities faster. Exploration mode stands out: users direct agents to probe resources in their accounts, generating visualizations or prototypes on the fly. Product teams get immediate feedback and testable ideas, all informed by prior discoveries. At the heart of it, we treat code as the core medium for these interactions—flexible for varied problems, easy to combine for bigger ones, and straightforward to learn from across examples. Our internal tests show the payoff: when agents face similar goals, shared memory raises success rates and lowers costs sharply, leaving single-agent or no-memory setups behind.

This approach draws from collective intelligence theories. Distributed systems thrive when experiences compound. We’re building on multi-agent frameworks, but as a protocol—not another model or tool. Protocols enable ecosystems: developers integrate easily via our OmniTool, which works with LangChain, CrewAI, AutoGen, and more. Enterprises gain reliable automation, communities contribute.

Why now? AI adoption is hitting reliability walls. Demos impress, but production demands consistency. Developer productivity is the hottest GenAI category in 2025, and agentic systems promise autonomy—if they can learn continuously. entourage addresses this by turning individual agent work into collective gains. We’re creating a system where intelligence scales with participation, solving real problems in coding, trading, and beyond.

At its core, entourage weaves the fabric for AI’s next evolution. True progress in AI won’t come from isolated models or bigger datasets alone. It requires experiential collective intelligence—shared memory that lets agents build on each other’s real-world discoveries. Without it, we stay stuck in silos. With it, we unlock systems that adapt, generalize, and improve indefinitely. If you’re in AI, consider joining this shift. Reach out to learn more about entourage.