Skip to content

The agent framework (v2 preview)

Preview

Framework v2 is not yet available to all accounts.

Framework v2 is the next version of the Anychat agent framework. It keeps the way you already describe an agent and changes what happens underneath, so that the agent's behavior is easier to predict, check, and explain. This section describes v2 for the person who runs the agent: a marketer, an operations lead, or a partner setting agents up for clients. It assumes no technical background.

What you give an agent

An agent is made of three things, and you supply each one differently.

You give the agent What it is What you do with it
Its Brief One document: who the agent is, how it sounds, what it never does, what it is trying to achieve, and a Playbook for each kind of situation. Write it.
Its Capabilities Things the agent can do beyond talking: book calls, answer from your documents and catalog, remember people, alert your team, call your systems. Switch them on and configure them.
Its Channels Where it speaks: RCS, Apple Messages for Business, WhatsApp, SMS, Webchat. Connect them.

Two more layers sit alongside, and both are described in this section: Flows, which give a Playbook a fixed path when the situation calls for one, and Policies, which the platform enforces no matter what the agent would otherwise say.

Turning v2 on for an agent

v2 is a preview, offered to orgs that hold the framework-v2 feature; ask your Anychat contact to enable it. Once it is on, every agent's Overview page carries a Switch to v2 preview control (and, for an agent already on v2, Switch back to Generic). The switch is per agent and reversible: the Definition is untouched, the change takes effect on the agent's next turn, and a conversation in progress continues in the new runtime's default Playbook with the same transcript. An agent on v2 shows a v2 preview badge next to its name. Before switching, look at the Readiness card: lint lists anything in the Definition that v2 reads differently or ignores (procedure requiredTools, the retired skills library, markup inside message templates). New agents can also be created on v2 directly.

The one promise

The agent only ever reads what you can see.

On every turn the agent works from three things: the Brief, each enabled capability's short "what this adds" text (shown read-only on the Capabilities page), and a one-line profile of the channel the conversation is on. Nothing else is added behind the scenes. If you want to know why the agent said something, you can open the Brief, read the capability's text, and look at the same channel line the agent saw. The console's Preview what the agent sees shows the composed whole for any Playbook and channel, and every real turn keeps a record you can open afterwards (see How a turn works).

The worked example

Every page in this section uses the same agent. Mia is the sales-development agent for Glow Aesthetics, a med spa. Mia answers questions about treatments, qualifies people who are considering a visit, books consultation calls with a member of the Glow team, and hands people to a human when they ask for one. Mia talks on RCS and falls back to SMS.

Mia's Brief has five Playbooks: Welcome (the default), Answer treatment questions, Qualify, Book a consultation call, and Talk to a person. Mia has four capabilities switched on: Scheduling, Knowledge, People, and Notifications.

One turn, in words

A person on RCS writes to Mia: "Do you do laser hair removal, and could I get a call this week?"

  1. Anychat checks the message first: this person has not opted out, and the channel allows a reply right now.
  2. Anychat picks the Playbook that fits. The message is about a treatment and about booking; Answer treatment questions matches first, and its when-to-use line says to hand to Book a consultation call once the person is ready.
  3. Anychat gathers what the agent reads this turn: the Brief (with the active Playbook's full body and only the names and when-to-use lines of the others), the Knowledge and Scheduling "what this adds" text, the RCS channel line, this person's record, and the conversation so far.
  4. The agent looks up laser hair removal in Glow's documents, then writes its reply: a short answer, and a question about whether the person would like to see available times.
  5. Anychat checks the reply against Policies, shapes it for RCS (a text bubble and two suggestion chips), and sends it.
  6. The turn is recorded: what was read, what was looked up, what was said, and what was sent.

The rest of this section takes each of these apart.

Pages in this section

  • The Brief: the document, section by section, and how the agent decides which Playbook is active.
  • Capabilities: what each one lets the agent do, what you configure, and what happens when it is not ready.
  • Playbooks and Flows: open-ended Playbooks, Playbooks with a fixed path, and when to use which.
  • Rules and safety: Guardrails, Policies, reviewed copy, and what the agent cannot do by construction.
  • How a turn works: the pipeline from an incoming message to a recorded turn.
  • Checking your agent: readiness, lint, preview, test conversations, and evaluations.
  • Glossary.