The things that make an AI agent reliably accurate at scale are the same things that help people succeed. As an early-stage startup, Mate builds agents that help security teams defend some of the largest companies in the world. Getting an agent to give the right answer once is easy, but getting it right a million times, across messy data, is far harder.
Building AI agents for large enterprises taught me what keeps them accurate in ambiguous, high-stakes situations. Agents need shared context, clear structure, distilled patterns they can rely on at scale, a filter for noise, and an understanding of how things connect. People need exactly the same things, which is why those lessons now shape how I lead, how I grow teams, and how we’ve laid the foundation for scaling Mate from day one.
Here are five lessons from our agents that shaped how we build our team.
1. Context becomes culture
An agent without context guesses, but with a shared context, it can build on what came before. With context, agents understand how a business operates before they investigate an alert. They can tell whether something suspicious actually warrants attention, reach evidence-based verdicts, and explain how they got there. Context is what turns a fast agent into one a security team can trust.
Teams work the same way. Context at work means knowing why past decisions were made, who owns what, and how each person’s work connects to the customer. With it, people can judge what deserves their attention, make decisions without waiting for sign-off, and explain their reasoning to anyone who asks. Context also means that everyone uses the same terms and knows exactly what they mean. Over time, that shared language turns into a shared culture.
For us, building an internal context layer starts with wide adoption of AI tools that help create the company’s shared knowledge base. We say yes to every tool and every experiment our team asks for. That way, everyone has reliable context so they can save time and move fast. Every new hire gets access to a company-wide context structure on GitHub on their first day, a shared foundation the whole team works on top of.
Shared context also compounds: every document, decision, and lesson someone adds makes the next person’s work easier, just as an agent gets sharper as its context grows richer.
2. Structure lets people reason
Agents can reason only over data that has been shaped and indexed to match how they think and what they understand.
People also need structure they can understand. Clear processes remove friction and free their energy for creative work and for what customers need. When ownership is clear, nobody burns a day figuring out who to ask, and when decisions are written down, anyone can find out why we did something without chasing the person who decided it. As a result, people who understand the structure move faster and deliver better results.
3. Scale requires compression
As an agent handles more, it needs a compressed model of the world to stay accurate, built from distilled patterns that hold up at volume.
Companies face the same math: we serve Fortune 500 companies, and every new customer expects the same standard as our first. As the team grows, we compress what we know into playbooks, shared frameworks, and clear quality bars, so knowledge is spread across the whole team. Someone who joined last month should be able to deliver the same level of expertise as someone who has been here since the beginning.
For an early-stage startup with limited resources, compression can start small. The second time someone solves the same problem, they write down how, and that becomes the first draft of a playbook. AI tools do much of the heavy lifting, turning call notes, Slack threads, and investigation write-ups into documentation in minutes. Each playbook has one owner and fits on a page, so it stays easy to update. Our agents distill patterns as those patterns repeat, and our playbooks grow the same way.
4. Filter the noise
Past a certain point, more data makes an agent worse, because the more volume there is, the harder it gets to tell what matters from what’s garbage.
Growing companies hit the same wall. With more customers, more channels, and more requests, everything starts to look urgent. Our technology filters out irrelevant data before an agent ever sees it, and we do the same for our people. Every week, leaders take on the filtering, so each team knows what matters now, what can wait, and where to put its attention.
Filtering also means making trade-offs. Every new priority competes with an existing one, so when something new comes in, leaders decide what moves down the list.
5. Meaning comes from connections
Putting all the logs in one place is a good start for an agent, but it also needs to understand what the data means and how one event connects to another. That kind of understanding comes from experience.
New team members face the same challenge: they can read every account file and still miss the bigger picture, so our onboarding is built around connections. We walk new hires through how different customers, accounts, and use cases relate to each other, so they learn how a pattern in one account can help another and how the pieces together create more value for customers. Our longest-tenured team members play a big role here. They have seen how patterns repeat across customers, and sharing that perspective turns individual experience into team knowledge.
The common thread
We put enormous effort into designing the right environment for our agents, and the people building them deserve at least as much thought. Give your people shared context, structure that fits how they work, a way to scale, clear focus, and a sense of meaning. Get those right, and both your agents and your people will do their best work.
Asaf Wiener is the CEO of Mate, the open foundation for agentic cyberdefense. Mate was selected to LinkedIn’s Top Startups list for 2026.
