What is an AI Employee

Expert in a specific role, like front desk, website engagement, inside sales etc. They know the vocabulary if the business, the workflows, the integrations and the pitfalls to stay away from. AI Employees are useful out of the box and can stand in for tasks or full roles.

How Twig builds AI Employees

Twig has a Agent Specification Language a proprietary language the lets you define agent behavior in a code like script. This is important as it makes AI agents deterministic. Which is important when you want to depend on them to take over tasks.

Agent Scripting Language

There's a difference between an AI feature and an AI employee, and it isn't the model. An AI feature answers a question and forgets you. An AI employee has a job. It works a process end to end, carries context across a conversation, knows when it's allowed to take an action and when it isn't, and can be held to a standard the way you'd hold a person to one.

Sera, our AI front desk agent, is built to be the second thing. The reason she can be is a layer we built called Agent Scripting Language. I'm not going to open the hood all the way here — but it's worth a glimpse, because it's the part most "AI agent" products are missing.

Language is easy, process is hard

Modern LLMs are extraordinary at language and mediocre at process. Hand a model a multi-step procedure in prose — greet them, capture what you need, qualify them, and only then take the action — and it will follow it most of the time. "Most of the time" is the problem. Deep in a real conversation, agents skip a step, re-ask for something they were already told, or confidently invent a fact under pressure.

Most teams reach for one of two tools, and both fail in their own way. The mega-prompt stuffs every instruction, edge case, and fact into one giant system prompt — it works until it's a wall of text where the rules contradict each other and no one can change one behavior without breaking three others. The fully autonomous agent hands the model a pile of tools and hopes it sequences them correctly — and you get improvisation where you wanted a procedure.

A front desk or an SDR doesn't need improvisation in its process. It needs to run the same reliable procedure every time, in language that still feels human.

The glimpse

Agent Scripting Language is how we close that gap. Instead of describing an employee's job in one long prompt and hoping, we structure it — so the model does what it's genuinely great at (understanding the message, choosing the words) while the system stays in control of what happens when.

The shape of it, without the internals: an employee's job is expressed as a set of small, composable behaviors rather than one monolith. Each one carries its own sense of when it should run and what has to already be true before it's allowed to. The result is an agent whose process is a property of the system, not a hope about the model — one you can reason about, change, and trust one piece at a time.

That last part is the quiet payoff. Because behavior is structured rather than stuffed into a prompt, the agent won't take an action it isn't allowed to take — booking before it has what it needs, escalating when it shouldn't — because the guardrail is structural, not a polite instruction it can drift past. And the behavior becomes testable: we can replay real conversations, confirm the agent still does the right thing, and catch drift before it ships — so a model swap or a small change doesn't quietly degrade the employee weeks before anyone notices.

Why it matters

AI features are easy to ship and easy to forget. AI employees have to show up and do the job the same way every time. That reliability isn't something you can prompt your way to — it has to be built. Agent Scripting Language is how we build it, and it's why Sera can work a front desk instead of just answering a question.

We'll share more over time. For now, this is the glimpse.