Michael Bajwa and Qian Zheng helped build a product now generating $1.2 billion a year. Then they left to build something new, a platform where the intelligence a company creates compounds for the company itself, not the software it runs on.
At Convergence, the two had built Proxy, one of the first AI agents that could navigate any software without API integrations, using long-term memory to work out what to do next. Michael, the first product hire, took it from zero to $1 million in ARR in ten weeks. Qian, the first engineering hire, built the production system from scratch. Salesforce acquired the company eleven months in, and the team went on to ship Agentforce, now at $1.2 billion in ARR.
That gave them a vantage point nobody else had. From inside Salesforce, they watched every company running Agentforce make Salesforce smarter, not themselves. Every agent interaction, every automated decision, every workflow was building institutional intelligence inside the vendor's infrastructure. The companies paying for it walked away with nothing to show for it.
So they left to fix it. Their thesis for Zaro was simple: a company's intelligence should compound for the company that builds it, not the vendor it rents tools from.
We are proud to be leading Zaro's $5.1 million pre-seed round, investing alongside the people who knew their work best - Marvin Purtorab and Andy Toulis, Convergence founders, and the co-founder of Hugging Face and the former CEO of GitHub.
One of the core reasons for our investment was speed. Michael and Qian are setting the bar for pace of development and showed us a working product three weeks after we invested.
In a category where a hundred tools promise the same thing, pace is the game. The teams that ship fastest will have a shot at defining the category.
The company is the context
Most agent systems work like a relay. Agent A does a task and hands the output to Agent B, who does the next task. Each agent only sees what was passed to them. If something useful comes up midway through a task, the next agent has no idea. Humans fill the gaps and every new task starts from zero context.
Zaro replaces the relay with a shared space that holds your company's knowledge. Every agent reads from and writes to it at once, like a document the whole team edits live instead of emailing versions back and forth. Your emails, call recordings, Slack threads, and documents all live there.
That is what removes the connectors. Most tools build a bridge from your CRM to your agent, then another from your email, then another for every system after that. Zaro inverts it. Instead of connecting tool to tool, it makes the company itself the thing agents plug into. There is no bridge to build, because everything is already in one place.
Once agents have shared memory, they can also act on it. In our use at Cherry, that meant asking for a research tool and getting one that already knew our portfolio. And because the context is shared but the surface is not fixed, everyone works with it their own way. Two people can use the same underlying app through completely different interfaces, rather than fighting over one shared view.
The longer Zaro runs, the more it understands about how your company works so it gets more useful over time and can eventually run familiar workflows without being asked.
The context problem is worth solving now
According to McKinsey, 62% of enterprises are experimenting with agents but only 23% are scaling them. The gap exists for the same reason our own agents kept failing. Most agents work great in isolation. But the moment real work kicks in, with all its messiness and history and context spread across different tools, the agents fall apart.
The longer enterprises run agents without shared memory, the more intelligence compounds on the wrong side. Every agent interaction that doesn't feed back into the company's own context is making someone else's platform smarter.
MCP is now the default standard for enterprise AI. Every major platform supports it. The infrastructure for shared context exists. The question is whether the intelligence agents generate stays with the company or with the vendor.
"Context compounds," as Qian puts it. "Models commoditise. The platform does not."
Proof of conviction
A few months after we invested, team Cherry were among the first to use the product. We had spent a year trying to solve the context problem with connectors and Zaro didn’t use a single one.
We started running real work through it. Every workflow we ran added to what the platform knew about how Cherry works. Now, we have replaced several internal tools, ran event research, and dropped 15 tool subscriptions we'd been paying for for years. The intelligence stayed with us, rather than the platforms we'd left behind.
Almost everyone who tries Zaro for ten minutes comes back. Everyone else was building better connectors between tools. Michael and Qian made the company itself the thing agents plug into. Nobody else was doing that.
