Skip to main content
Everyone’s first question about voice AI is the same: can I trust it on a real call, with a real customer, when I’m not listening? It’s the right question. A voice agent that improvises at the wrong moment costs you the lead. Our answer is skills. You already know the idea from Claude Code and Codex: focused capabilities the model reaches for at the right moment, instead of one giant prompt holding everything at once. Your voice agent carries them too.

​
A new way to think about a voice agent

For years a voice agent was one long script. Every rule and objection packed into a single wall of text, with the agent trying to hold all of it at once. It works until it doesn’t. The more you add, the more each instruction gets buried, and the more the agent drifts. Skills flip that. The agent carries a set of small, focused plays and reaches for the right one exactly when the moment comes up. It’s closer to how a good rep actually works: they don’t recite everything they know on every call, they pull the right move when the call calls for it.

​
Why it makes calls more reliable

  • Progressive disclosure. The agent sees the triggers, not the full text. It pulls one skill’s instructions the instant that moment lands and ignores the rest. More skills makes it sharper, not foggier.
  • Fewer hallucinations. Each moment gets its own tight brief instead of being lost in a mega prompt, so there’s far less room to improvise.
  • More deterministic. For the things that must be exactly right, do we cover this ZIP, transfer here, tag this lead, a tool runs real code and returns a fact instead of a guess.
  • Always improvable. Every skill is a small thing you can watch, edit, and sharpen on its own. Your agent keeps getting more reliable because you tune specific plays, not a wall of text.

​
Two kinds of ability

Skills you write

Rules in plain English: when this comes up, do that. No code. Your judgment, taught to the agent.

Tools we ship

Built abilities you add and configure, deterministic where it counts. A tool checks a ZIP, transfers a call, tags a lead. You don’t build tools; we maintain them so they work every call.
In the dashboard both live in one panel, Skills & Tools: everything this agent can do during a call. It’s all per agent, and nothing here makes the agent dial. That’s still the campaign. Appointment Booking, Live Transfer and Add a Tag are capabilities. You’ll find them in the same library as the other tools. Each capability brings its in-call action and the related settings together, such as the booking calendar or transfer destination. An ability being available in the library does not mean it is installed; you choose which ones your calls need.

​
During the call, or after it?

This is the first question to settle, because it decides whether you want a skill at all. A skill runs during the call, while the caller is on the line. A follow-up action runs after the call ends, off the outcome the agent recorded. They happen at different moments, and each one is impossible in the other’s. During the call: Live Transfer, Appointment Booking and Service Area Check. After the call ends: follow-up actions such as Add tag, Send SMS and Move stage. During the call: Live Transfer, Appointment Booking and Service Area Check. After the call ends: follow-up actions such as Add tag, Send SMS and Move stage. After the call is the default. Something belongs during the call only if the call actually depends on it:
  • Transferring to a person. The whole point is connecting someone now.
  • Booking, when the caller has to hear the time and agree to it.
  • Anything the agent needs an answer to before it knows what to say next: a service-area check, a tag lookup that changes the conversation.
  • A label whose only job is to fire something the caller is being told about right now. “I’m texting you that link” is real work happening on the call, and they’ll notice if it doesn’t arrive.
Everything else is follow-up: labels for what happened (booked, transferred, qualified, not interested), notes, pipeline moves, follow-up texts and emails. Nobody on the call perceives any of it.
Why after the call wins whenever it’s allowed. It runs every single time that outcome happens, instead of only when the agent decides to reach for a tool. It runs off what actually happened on the whole call, not a guess made halfway through. It can branch on a value. It doesn’t spend the one slot each ability gets. It costs the call no waiting. And you can change it later without touching the script.
A connected transfer ends the agent’s turn. Once the caller is handed to a person, the agent is off the call and can’t do anything else, so “tag everyone who gets transferred” can never be a mid-call ability, no matter how it’s worded. That one goes on the campaign’s follow-up, on the Converted outcome, which is where a connected transfer lands when Live Transfer is installed. The same is true of anything meant to happen once the call is over.
None of this makes mid-call tagging a lesser option. It’s exactly right when you want your own automation firing while the caller is still listening. It just isn’t the place for a label that only records what happened.

​
Any ability installs on any campaign

There is no compatibility rule between abilities and campaign types. A campaign carries a set of abilities and installing one adds to that set, so an appointment booking agent can absolutely also transfer to a human. Nothing needs a separate campaign for that. What is fixed is the campaign’s trigger: what starts it, a tag landing or an inbound call or a booked appointment. That is chosen once, at creation. “Also transfer to a human” is an install. “Also answer inbound calls” is a different campaign.
Installing an ability moves the go-live bar. Readiness follows the abilities the agent actually has, not the campaign type. A live campaign keeps running after an install, but the next time it goes live it is blocked until the new requirement is met. The same checks run before a test call.
So the install is only half the job. Whatever it newly requires has to be set in the same sitting, or the agent has an ability it cannot use and the campaign cannot go live. For a complete collection change, read Editing several skills. On an agent tagged TFU Live the same idea is arranged differently: abilities are equipped on a canvas, and anything that must happen every time one succeeds is a step in that workflow rather than a line in the script. See Capabilities and workflows.

​
In the API

A skill is one record on the agent: The set is versioned. Reads return version; send it back as expectedVersion on a write, and a stale one gets 409. See Conventions. Three more reads help you manage them:
  • The library lists everything this agent could carry, each with kind (skill or tool), a status of available, added_as_skill or installed_tool, canAdd, and the fields you would configure.
  • The digest is one line per skill: what it is for and whether it is on.
  • The tools inventory lists the tools installed on the agent with a freshness of live or cached, and an editableDescription you can change to adjust when the agent reaches for that tool.
Endpoints: