How the agent behaves, in instructions.md, and the named procedures it
follows, in workflows/. A workflow has a name, a one-line When to trigger, and a
Procedure.
Build an agent
Everything about how an agent works lives in its Build screen. Edits collect in a draft, and Publish turns the draft into the next version.
The Build screen
Open it with Edit on an agent. Five tabs hold the agent. On a phone they become a Section menu. The tag on each tells you when a change reaches the agent's sessions.
Specinstructions.md · workflows/waits for Publish
Knowledgefiles it readswaits for Publishown files apply at once
The shared knowledge bases it reads, then Files only this agent has, up to 25 MB each. Put PDFs in a knowledge base: only a knowledge base turns them into text the agent can read. Adding a knowledge base waits for Publish; the agent's own files apply at once. See Knowledge.
Connectionsapps · databases · toolswaits for Publish
The apps it uses and whose account it signs in with, its databases, and any MCP servers or HTTP tools. See Connect apps and data.
Secretskeys and passwordsvalues apply at once
Values the agent's tools need, such as an API key. The name is written in UPPER_SNAKE_CASE. The value is write-only: nobody can read it back, you included. See Secrets.
Triggerschedule · webhookapplies at once
Run it on a schedule (Every 15 minutes, Every hour, Daily at 9:00, Monday at 9:00, or Custom cron) or from a webhook. Saving makes a new version straight away, without Publish.
Guardrailsstay on topic · ruleswaits for Publish
What the agent must never do, enforced even when someone tries to talk it out of it. Stay on topic politely turns away requests unrelated to its job; turn it off for a general assistant. Each rule is one plain-English sentence, up to 20 of them. Every rule is checked on what people send, what the agent tries to do and what it says back. See Guardrails.
The Agent Builder
The Agent Builder is the assistant beside the Build screen. Tell it what to change in plain language. It edits the draft for you, and you still decide when to publish.
| It can | It cannot |
|---|---|
| Rewrite the instructions | Set a schedule or a webhook: use the Trigger tab |
| Add, change or remove workflows | Publish: you press Publish |
| Add, change or remove connections, after asking whose account to use | Store a secret's value: type it in Secrets |
| Name the secrets the agent needs | Change the grade or the agent's name |
| Set questions the agent asks at the start of a session | Add files: use Knowledge |
Messages to it are free: they use none of your credit. Reset starts a fresh conversation with it.
Draft, Publish and versions
You edit
The draft
Instructions, workflows, connections, knowledge bases and the grade. The agent reads
"Draft · unpublished changes". Nothing running sees the draft.
↓Publish
AgentNava
The next version
The dialog shows "Publishing saves these edits as v3" and checks first: "Everything this agent
needs is attached", or a list of what is missing.
↓
What moves
New work uses v3
New sessions and scheduled runs use v3 at once. A person's conversation moves to v3 the next
time it resumes. Sessions already running keep their version. The website chat moves to v3 too, unless v3 needs
each person's own account or a secret.
Some changes apply at once
Saving the Trigger tab, renaming the agent, its own files, secret values and app sign-ins all reach the agent straight away, without Publish. Treat them as changes to a live agent.
There is no undo yet
A draft cannot be thrown away with one button, and there is no way to roll back to an earlier version. To undo an edit, change it back in the draft, then publish again.
Grades
A grade sets the model an agent runs and the price of each turn. Change it with the Grade chip in the Build screen. The change waits for Publish, like any other edit.
| Grade | Per turn | For |
|---|---|---|
| Lite | 0.01 credits | Everyday work at the lowest price: reading, writing, routine reasoning. |
| Standard | 0.02 credits | The fastest model we run, for everyday work and live conversation. Every new agent starts here. |
| Premium | 0.05 credits | The hardest reasoning, for work that has to be right. |
| Best | 0.2 credits | The strongest model we run, when quality beats everything. |
A credit is a dollar. Credits and billing shows what work costs in practice.
Answer with widgets
An agent can answer with cards, tables and charts inside its replies, where the chat can show them, such as the website chat. It is on for every new agent. Turn it off with the Widgets switch under Version in the Spec tab, or ask the Agent Builder to answer in text only, for an agent such as a legal assistant or a voice agent. The change waits for Publish, like any other edit.
Test it
Test, on the agent's card and on its page, starts a session with you as the user. It runs the published version with its real tools, and each turn is charged at the agent's grade. The session then appears in the agent's Sessions list.
So publish first, then test. A test straight after an unpublished edit runs the old version. If the agent asks questions at the start of a session, a New session dialog asks them first.
Pause, resume and delete
Pause, on the agent's page, stops all new work on every channel: chat, website, API, webhooks and schedules. A turn already running finishes. Resume lets it take new work again.
To delete an agent, open the ⋯ menu on its card and choose Delete. Deleting is permanent.