Agents
Instructions, documents, skills, memory, models, publishing, and how an agent hands a conversation to a person.
An agent is a job description plus what it is allowed to know. You write what it is for and how it should behave, give it documents to answer from, and choose a model. Kumoda runs it whenever a customer writes on a channel you connected.
Starting points
Open Agents to see your agents, or Playbooks for ready-made ones. A playbook opens the agent builder with the agent already filled in, and you can edit anything. The starting points are:
| Playbook | What it does, as the app describes it |
|---|---|
| WhatsApp enquiry responder | Answers common service questions from your documents and asks for missing details. |
| Follow-up drafter | You paste an enquiry; it drafts the reply and the next follow-up. |
| Payment reminder drafter | You paste an overdue invoice; it drafts a polite reminder, firmer each time. |
| Support reply drafter | You paste a customer message; it drafts a clear, kind reply. |
| Handover assistant | You describe how you do something; it turns it into steps your team can follow. |
| Writing partner | Drafts and tightens your messages, posts and emails in your voice. |
| Offer sharpener | Pushes on your offer: who it is for, the promise, the price, what is unclear. |
| Start from blank | Write your own instructions. |
The Quickstart uses the first one.
The builder
The builder has five steps: Purpose, Identity, Knowledge, Tools and Publish. Changes are saved as a draft as you go.
Purpose
One or two sentences on what the agent is for. It helps you and your team remember its job, and Flow and your other agents read it to know when to ask this agent (see Flow).
Identity: names, model and instructions
- Internal name: only your team sees it.
- Customer-facing name: shown at the top of the chat window. Optional.
- Opening message: what the agent says first.
- Instructions: up to 4,000 characters. Write who the agent is, what it answers, and what it always does.
- Rules it always follows: short rules, up to 12, for example "Never state a refund amount". Rules are for the things that must never go wrong.
Kumoda also tells every agent to treat the results of tools as data, not as instructions.
Models
Choose a provider, then a model. Each model card shows bars for speed, quality and cost (for cost, more bars means pricier). The Anthropic models offered are:
| Model | Described in the app as |
|---|---|
| Haiku 5.5 | Fast and light. Good for simple, repeated questions. |
| Sonnet 5.5 | The all-round choice for most agents. Marked as recommended. |
| Opus 5.5 | The strongest. For hard judgement, and the priciest. |
Which models you can pick depends on your plan: the builder shows what your plan includes. Other providers are listed too, and their models can be picked once Kumoda has that provider connected. You can change the model at any time, and lighter models use fewer credits. See Plans, credits and billing.
Knowledge: documents
Knowledge sources are what the agent is allowed to say. It searches them for factual questions and answers in its own words. If nothing matches, it says it does not have that to hand. Without any sources, the agent answers from its instructions alone.
You can add a source three ways:
- Paste text: a policy, FAQ or price list, up to about 200,000 characters.
- Web address: a public page or PDF. The address must start with
https://. You can have Kumoda read it again later to pick up changes. - File: PDF, Word (
.docx), text, Markdown, CSV, JSON or HTML, up to 4 MB. A scanned PDF (pictures of pages) has no text to read.
You can edit the text of a source you pasted, and the agent answers from the new text straight away. Removing a source means the agent no longer answers from it.
Search is by shared words. Finding a passage that says the same thing in other words needs a setting on the Kumoda platform, which is shown in the app when it is not on yet.
Knowledge: skills
A skill is a procedure for a particular job, such as handling a refund or a booking. The agent reads only each skill's name and when to use it, and opens the steps when a question calls for them.
Each skill has a name, when to use it (one sentence, up to 300 characters) and steps (up to 8,000 characters). An agent can have up to 10 skills.
Knowledge: memory
Memory is what the agent keeps about each person it talks to, such as a customer on WhatsApp. Each person has their own memory, and it is never shared between people or agents.
- Recent messages: how much of the conversation the agent sees: nothing, or the last 5, 10 (the usual), 20 or 30 messages.
- Keep a note about the person: the agent writes down their name, what they want and what was promised, and reads it next time. This uses a little more of your credits on each message.
- Find older messages by meaning: lets the agent find something said long ago. It is available only when the platform has it switched on.
If long-term memory is not switched on for the platform, the agent sees only the recent messages it is handed, and the app says so.
Knowledge: helpers and message screening
- Helpers: other agents in the same workspace this one may ask for help with part of a question, up to three. What a helper uses is billed to your credits like any answer.
- Screen messages: checks each message for attempts to talk the agent out of its rules. A message that tells the agent to ignore its instructions or reveal them is not answered, and the person gets a polite refusal. It adds a small check to every message (about a tenth of a cent of your credits). If the check itself cannot run, the message is still answered, so a problem with it never leaves a customer without a reply.
Tools
Tools let an agent look something up or do something. See Tools and connections.
Test before you publish
Ask the agent what a real customer would, in the preview on the builder, and read the answers and the steps it took. A preview is a real model call and uses your credits. Nothing is sent to anyone.
You can also save questions as Tests, with what a good answer does. Run them against the saved draft before you publish a change. A member can keep up to five tests per agent, and each run uses credits.
Publish, pause and versions
The last step checks that the agent has a name, written instructions, a test in the preview, and at least one rule. Go live (or Publish version N after the first time) makes that version the one other systems get when they call the agent. It does not message anyone.
After that you can keep editing. Changes are tried in the preview and reach other systems only when you publish again. Each publish is a numbered version with an optional note, and Restore puts an older version back as a new one, so nothing is lost.
On the agent's page you can Pause and Resume it. A draft agent has Turn on. An agent's status (draft, live or paused) is separate from its channels: to stop customers being answered on a channel, use the Answer on switch for that channel under Channels, which pauses it. Deleting an agent removes the agent and its versions, and its past runs stay in the list.
The agent's page also shows its last seven days (runs, success rate, typical time, estimated cost), its recent runs, and Run automatically, which asks it the same question on a timetable. Anything an agent wants to change waits for your approval.
Handover to a person
On a customer chat (WhatsApp, the website chat window or the API) every agent has one tool, request_a_person. The agent calls it when the customer asks for a person, is upset, says the matter is urgent, or when it cannot help or is not sure. When it does:
- The conversation becomes yours. The agent stays out of it until you hand it back.
- The reason is recorded, and shown in Conversations as "Needs you".
- You are emailed, and so is everyone else who works the workspace (its Admins and Members), unless each has turned handover alerts off in their own profile. If you have named who handles the channel, only they are alerted (and you, when nobody has taken the chat), and if the conversation is already assigned to someone, the alert goes to that person.
- The agent tells the customer plainly that a person will reply.
In Conversations you can Take over at any time (without waiting for a handover), answer as yourself, Hand back to agent, Resolve and Reopen. A Viewer can read a conversation but not answer it. See Team and workspaces.
Two more things to know:
- On the WhatsApp quick start, if you type in a chat yourself, the agent stays quiet in that chat for eight hours (or until you hand it back).
- On the official WhatsApp connection, Meta does not tell Kumoda when you reply from your own phone, so the agent cannot step back by itself. Pause the channel, or take over in Conversations.
When your credits run out, a customer who writes gets one short note that a person will reply, and the conversation is handed to you. See Plans, credits and billing.