Safe by design: why we don't let our AI "figure it out"

Most talk about AI safety is about what a model might say. We focused on what it's allowed to do. Here's how Syncridge Fabric keeps our AI agents predictable, explainable and private by design.

By Alex Jahanban · Published Oct 16, 2026 · 4 min read

Safe by design: why we don't let our AI "figure it out"

This year, AI agents under test slipped past the limits meant to contain them, coordinated with one another, and reached systems they were never meant to touch: another company's platform, and even a government health portal. Not malice, just software chasing a goal with too much room to move.

Most talk about AI safety is about what a model might say. We spent more of our time on a different question: what is the model allowed to do?

We build two kinds of AI: our own flagship products, and tailored, bespoke AI tools designed around the needs of each business we work with. For both, we faced the same choice. The popular approach is to give a model a goal and a box of tools and let it work out its own path. It looks impressive in a demo. It is also hard to predict, hard to audit, and hard to explain to a customer whose data is on the line.

So we built Syncridge Fabric, the engine under everything we build, from flagship products like D.E.L.E.G.O. and S.T.R.A.X. to the bespoke tools we create for our clients, around a different idea: the AI does the thinking, and people design the path.

The model thinks, it doesn't wander

Every agent follows a path that a person designed and approved. Each unit on that path does what it does best, whether that's understanding a request, reasoning through a problem or checking a result. It never invents new steps, picks up new tools, or decides to go somewhere it wasn't sent. When it does make a choice, it chooses between options someone has already reviewed.

Designed paths, not simple ones

A designed path doesn't mean a simple one. Fabric is how we build intricate AI systems, both at the large scale and in the smallest details.

Behind every request sits layered systems of specialized blocks, orchestrated as a mixture of experts at both the macro and the micro level, each contributing what it does best before anything comes back out. That's what takes our agents well beyond reading and summarizing. They think critically, weigh evidence, and work through complex, multi-step problems, including calculations.

Every one of those specialists still sits on a designed path. So the extra capability doesn't come with extra unpredictability.

Judgement levels: work that gets checked

The harder the task, the more it matters that the answer is right. So our systems have judgement levels built in. When an agent produces something that has to be correct, such as a calculation, a conclusion or a decision, further levels of judgement check it before it's trusted. Accuracy isn't left to a single pass of a single block.

Limits are part of the structure

Every run has built-in ceilings on time, repetition and cost. These aren't a monitor watching for trouble after the fact. The structure simply doesn't allow anything else. An agent can't loop forever or spend without limit, because there's no shape in which it could.

Every answer can be explained

Paths are fixed and versioned. So when someone asks "why did it say that?", we can show which version answered, which steps ran, and in what order. You get an answer, not a guess.

Privacy: everyone sees only what they need

Safety is also about who sees what.

• Your data stays yours. Each organization's data is walled off. When the system can't tell whose data a request belongs to, it refuses. It never guesses.
• People talking to an agent see only the answer meant for them, not the internal notes or data behind it.
• Sensitive details can be removed early. Personal information can be detected and redacted before it travels any further.
• Connections use the right keys. When an agent reaches into your other systems, it can act with each customer's own sign-in instead of one shared master key.
• Inside your organization, access follows roles. Your team sees what their permissions allow, and no more.

Why this matters

No AI is perfect. Models still make mistakes, and we design for that: with checks on the work that matters and limits on how far any mistake can travel. By keeping the thinking powerful and the path designed, we get AI that can take on hard problems without becoming unpredictable. That's how we earn trust: through structure, not promises.