FLAE

English

Tiếng Việt

English

Tiếng Việt

Appearance

Chat with the Assistant ​

What it is ​

The FLAE Assistant works inside one Workspace. It explains how FLAE works, answers questions about your Agents, Tools and connected Apps, reads your live App data, and builds draft Tools, Skills, Agents and Workforces for you to review and publish. Find it in the left menu under Chat, which is also the page FLAE opens when you enter a Workspace.

The Chat page also shows Needs your decision (runs waiting for a person, with Open Agent Inbox) and Running (runs in progress, with See all runs).

Key concepts ​

  • Chat — one conversation with the Assistant, titled by your first message. Your chats are private: other members cannot see them. Chat history lists them by Today, Yesterday and Earlier, with Search your chats and the filters All, Questions and Build. A chat becomes Build once the Assistant shows a plan or changes something.
  • How it works on a request — it picks one of three ways:
    • Ask — a question. It looks things up and answers; it changes nothing.
    • Small edit — one Tool, Skill, Agent or Workforce to change, or one Tool, Skill or Agent to create. It does it straight away.
    • Build — two or more objects, or a new Workforce. It shows a plan first.
  • The product guide — to explain how FLAE works or what an App's actions do, the Assistant reads this guide (you see a step labelled docs.read in Steps the Assistant took). Ask it "how does …" questions the same way you would read these pages.
  • Steps — Steps the Assistant took lists each lookup, build and test, with its status and time. Click a step to see its Input and Output. A step marked review is a check of earlier work.
  • Plan card — Plan: the goal, each task (Tool, Skill, Agent or Workforce; Create or Update and its name; "Done when: …"), and You need to do: the steps only you can take. Press Approve or Request changes.
  • Assistant tasks — after you approve, the task list above the reply box shows each task as Pending, In progress, In review, Fixing (with "Fix 1/3" …), Done or Failed.
  • Approval card — before the Assistant changes data in a connected App (sends a message, creates an order), it shows the action and its exact input. Nothing runs until you press Approve; Reject stops it.
  • Changed — at the end of a turn, the list of Tools, Skills, Agents and Workforces actually saved in that turn, each Created or Updated and linked to its builder, with "Publish in this order: Tool → Skill → Agent → Workforce".

How it works ​

Asking

  1. Type your question on Chat and press Send. A new chat opens.
  2. The Assistant looks things up: this guide, your Workspace's objects, the App catalog, and, for questions about your own data (orders, customers, messages), read actions on your connected Apps. Reads need no approval.
  3. It answers in the language of your latest message.

Building

  1. It looks first. It lists what the Workspace already has and reuses fitting Tools, Skills and Agents instead of making copies.
  2. It may ask you a short round of questions: where requests come in, where the facts live, what the Agent may do alone, when to hand over to a person, who approves writes.
  3. Plan card. You approve it, or press Request changes and say what to change; it then shows a new plan. Typing a reply while a plan waits counts as Request changes.
  4. It runs through the plan without stopping to ask: tasks that others depend on first, then Tools, Skills, Agents and Workforces. Each task is built, then checked by a separate reviewer that can only read and test. A task that fails review gets up to 3 fix rounds; after that it is Failed and the Assistant moves on.
  5. Final check and report: what it created and changed, the order to publish in, any failed task, and what you must do yourself.

Testing. The Assistant tests only with dry runs. A Tool test simulates reads and writes with sample data. An Agent test (chat Agents only) sends 1 to 3 sample messages; reads hit your real Apps but writes are simulated, and the test stops at the first Tool that needs approval. It never runs a Tool for real.

Publishing. The Assistant publishes a Tool only after a passing test on that exact draft, and never one that a published Agent or Workforce already uses. It never publishes a Skill, Agent or Workforce: you do that in the app, Tool → Skill → Agent → Workforce.

Approval of write Tools. A write Tool the Assistant puts on an Agent is always set to ask a person first. Only you can switch it to Auto, in the Agent builder. On a Workforce, it sets connections into write Tools to Needs approval and leaves the Approver for you to pick.

Knowledge. It can list your Knowledge Bases and attach them to an Agent. Uploading documents happens on the Knowledge page.

Common tasks ​

Ask how something works

Ask, for example, "How do approvals work?" or "What can the Haravan App do?". The Assistant reads the matching guide page and answers from it.

Ask about your data

Ask, for example, "How many Haravan orders came in today?". It runs read actions on the connected App and answers from what they return. An App that is not connected: it tells you to connect it on the Integrations page.

Make a change in an App

Ask for the change ("Send this customer a Zalo message saying …"). Check the approval card, then press Approve, or Reject and give a reason. Typing a reply while the card waits rejects it, with your message as the reason. The Assistant does not try a rejected action again; it asks what you want instead.

Have the Assistant build an Agent

  1. Describe the job: who it serves, on which App, what it may do, when to hand over.
  2. Answer its questions, then approve the plan.
  3. When it finishes, open each item in Changed, review it, and publish in the order shown.
  4. Do what You need to do lists: connect an App, pick approvers, publish, set up the Automation that puts the Agent to work (see Automations).

Continue or find a chat

Open Chat history, search or filter, and click a chat. Write under Reply to ask a follow-up or ask for a change. New chat starts a fresh one.

Stop the Assistant

Press Stop while it works. It stops before its next step and shows "Stopped.". Work it already saved stays saved.

Limits and rules ​

  • Who can do what: everyone in the Workspace can chat. Workspace Admins and Builders (and Organization Owners and Admins) can have it build, test, publish Tools and read or change App data. For Operators and Viewers it can only look things up and read this guide; it says when your role cannot make a change.
  • The Assistant cannot: publish Skills, Agents or Workforces; turn a write Tool to Auto; create Automations or bind a Facebook Page; connect Apps; upload Knowledge documents; create voice Agents; run Evals; delete or archive anything.
  • Answers about your data come only from what the App returned. Each write needs your approval on its card, every time.
  • One turn at a time per chat: the reply box is locked while the Assistant works. A message can be up to 8,000 characters.
  • A turn that runs too long ends with "The Assistant stopped at its run time limit."; one that fails shows "The Assistant could not finish this turn.". Run time is the only limit a turn has today. Chats from before November 2026 may end with "The Assistant stopped at its cost limit."; that limit no longer exists.
  • The Attach a file (paperclip), Call an agent or tool and Agent: Automatic buttons under the message box do nothing yet.
  • Ask Assistant on the Automations page opens Chat. The Ask FLAE Assistant tab when creating a Workforce shows a sample only; ask in Chat instead.
  • Nothing can be changed in an archived Workspace.
  • Credits are not live yet, so they do not limit the Assistant, and a turn has no cost limit (see Account and billing).