Creating Agents
Build a new agent by talking to the Agent Builder
Agents are created and edited through a conversation with the Agent Builder. There's no upfront form — you describe what you want, the Builder creates it, and you keep iterating.
Start a new agent
- Open the agent selector and pick Agent Builder.
- Describe the agent in plain language — what it should help with, who uses it, anything specific it should know.
- The Builder asks for the minimum it needs (a name, a base agent, and which Kind if the intent is unclear) and creates the agent immediately.
- From there, you keep talking. The Builder edits instructions, greeting, capabilities, and the welcome screen as you describe changes. A live preview panel on the side shows everything taking shape.
You never have to fill in a blank form. Every field below ends up filled either by the Builder during the conversation or by you tweaking it directly in the preview.
Agent or Process — chosen from your first request
Every user-defined entry is either a runnable Agent or a documented-only Process (no Execute, no capabilities, hidden from selectors). The Builder chooses one based on how you phrase the request:
- Wording like "diagram", "draw", "flowchart", "visualize", "sketch", "map out", "SOP", "procedure", "checklist", "playbook", "runbook", "steps to…", "document how…" → created as a Process.
- Wording like "assistant that…", "agent that can…", "bot to…", or any list of capabilities, integrations, or actions → created as an Agent.
- If the intent is ambiguous, the Builder asks one short question ("Agent or Process?") before creating.
You can flip the Kind later from the agent's Properties. Promoting a Process to an Agent then lets you pick a base agent and turn on capabilities.
The fields you'll end up with
Required
| Field | What it is |
|---|---|
| Name | Display name shown in selectors |
| Identifier | Unique ID used for @mentions (e.g. sales_helper) |
| Base agent | The platform agent yours extends |
Identifier rules — starts with a letter or underscore; contains only letters, numbers, underscores, or hyphens; unique across your agents. Examples: my_helper, sales_q4, support-tier1.
Choosing a base agent — the base agent decides which capabilities are available, the reasoning ceiling, and whether an instance connection is required.
| If you need… | Pick |
|---|---|
| Quick data lookups, product knowledge | Advisor |
| Daily business operations, data management | Business Specialist |
| Customer-facing tasks, orders | Sales Executive |
| Complex configuration, custom logic | System Architect |
| Direct AI with no tools | Pure AI |
Optional
Reasoning level
Override the base agent's default. Available levels are capped by what the base agent supports.
| Level | Best for | AI cost |
|---|---|---|
| Fast | Quick lookups, simple queries | Lowest |
| Smart | Data analysis, business operations | Medium |
| Expert | Multi-step problem-solving | Higher |
| Genius | Maximum reasoning with extended thinking | Highest |
Leave it on Auto unless the agent consistently needs more (or less) than the base agent's default.
Every base agent has its own minimum, default and maximum — for example a base built for analysis won't let you drop below Smart, while a lookup-oriented base won't go above Fast. Anything outside that range is rejected.
You can also handle this in chat: ask the builder "what reasoning level is this agent using?" and it will tell you the level it actually runs at (and whether that comes from the base default or an explicit setting). Say "set the reasoning to expert" to change it — if the level is outside the base agent's range, the builder explains the range and offers the closest valid option.
Capabilities and approval modes
Each capability on the Capabilities tab has three settings:
| Mode | What it means |
|---|---|
| Off | The agent has no tools from this capability. |
| Ask | The agent can use the tools, but every sensitive action (sending an email, creating or changing a record, changing a document state) waits for your approval in the chat. This is the default. |
| Auto | Sensitive actions run immediately, without asking. |
You can set all three by talking to the Agent Builder, for example:
- "Turn on email and calendar, but ask me before sending anything." → both set to Ask.
- "Stop asking me before it books meetings." → calendar set to Auto, email stays on Ask.
- "It runs on a schedule at night — it must not wait for me." → the capabilities it needs are set to Auto.
Auto matters most for agents started by a schedule, a webhook, or an incoming email: with Ask, such a run simply stops and waits until someone opens the conversation and approves. The Builder will confirm with you before switching anything to Auto, and tells you whenever it changes an approval mode.
Instructions
Private guidance injected into every conversation. Users don't see it in chat.
Good instructions are specific, actionable, and focused on your particular use case.
Example:
You are helping with our Q4 sales pipeline. Focus on:
- Prioritizing deals over $50K
- Flagging stalled opportunities (no activity in 14+ days)
- Always check customer payment history before recommending discounts
When analyzing deals, always show the close probability and days in current stage.
Avoid: generic guidance the AI already follows, overly long instructions (focus beats volume), and anything that contradicts the base agent's purpose.
Greeting
A custom welcome message shown at the start of every new conversation with this agent. Use it to explain what the agent specializes in, suggest how to get started, and set expectations.
Example:
👋 I'm your Sales Pipeline Assistant, focused on Q4 deals.
I can help you:
- Review high-value opportunities
- Identify stalled deals needing attention
- Check customer history before offering discounts
What would you like to explore?
Suggested actions
Quick-action buttons that appear as cards under the greeting on the home screen when this agent is selected. They give immediate access to common tasks without typing.
Each action has:
| Property | Description |
|---|---|
| Label | Button text — keep it short |
| Type | Action (sends the message), Draft (populates the input), or Link (opens a URL) |
| Value | The message to send or URL to open |
Example for a Sales Pipeline agent:
- "Show high-value deals" (Action)
- "Stalled opportunities" (Action)
- "Open Sales Dashboard" (Link)
Welcome screen (WYSIWYG)
A what-you-see-is-what-you-get editor for the screen end users land on when they open your agent. It shows the agent's icon, name and description, the greeting message, the suggested-action cards, and a preview-only chat input — all live, all editable inline. Edits save automatically and update the same fields used everywhere else.
Shared agents are read-only in the welcome screen editor unless you own them.
Editing later
Just keep talking to the Agent Builder — tell it what to change ("shorten the instructions", "add a button that opens the sales dashboard", "switch the base agent to System Architect") and it updates the agent.
You can also edit any field directly in the side panel: greeting text, suggested actions, the welcome screen, capabilities, base agent, reasoning level, description, icon. Both flows save into the same agent.
Activate an agent from your email
You can ask the Agent Builder, in plain language, to make an agent react to incoming mail. It finds the right mailbox, creates the activation trigger, and turns the rest of your sentence into the note the agent receives when it wakes up.
Your own inbox
"Follow my email and notify me whenever a customer complains about a delivery."
The builder picks the mailbox you connected for yourself and creates an activation that runs on every new message.
A shared or team mailbox
"Watch the support@acme.com mailbox and draft a first reply for every new request."
Shared and delegate mailboxes work the same way, as long as that mailbox is connected. The builder tells you which mailbox it used, so you can confirm it picked the right one.
Only some emails
"Follow my email, but only messages from finance@supplier.com with 'invoice' in the subject."
Here the builder adds a filter to the activation, so the agent is woken up only for matching messages instead of every mail that arrives.
No mailbox connected yet?
If there's no mailbox to follow, the builder opens the email setup form for you right in the chat. Connect your mailbox (or a shared one), tell the builder you're done, and it continues with the activation.
Two things worth knowing:
- If mail syncing for that mailbox is stopped, the activation never fires — the builder points this out and offers to reopen the email setup so you can start it again.
- Nobody is watching the chat when an email wakes the agent up. If the agent should send replies or create records on its own, set those capabilities to Auto, otherwise the run waits forever for an approval nobody will give.
Ask the builder why a run behaved that way
The Agent Builder can read your agent's past runs, so you can debug in plain language instead of digging through history yourself.
Ask things like:
- "Check why it didn't send the email on the last activation."
- "Show me the last 5 activation runs — any failures?"
- "The agent created the deal but stopped. What happened?"
- "Why did yesterday's scheduled run fail?"
The builder looks up the run, reads which steps and tools actually ran, and tells you the concrete reason — for example a tool error, a missing capability, or a run that paused waiting for your approval before a sensitive action such as sending an email. It then proposes a fix (an instruction change, an extra capability, a trigger setting) and can apply it right away.
This is read-only inspection: the builder can explain a run but cannot approve, cancel or re-run it. To act on a paused run, open it from the Conversations tab.
The Executions tab
The same information is available visually. In the agent's workspace, open the More menu → Executions to see every real run of the agent, newest first:
- When it ran, whether it was a chat, an activation (with the trigger name) or a delegation.
- Its status — completed, failed, running, or awaiting your approval — and the failure reason when it failed.
- Click a run to see its full timeline: your messages, the agent's replies, each tool it called with the arguments and the result or error, plus any ERP.net request failures.
Design conversations with the Agent Builder stay in the Conversations tab; Executions shows only real runs. Like the builder's own inspection, this view is read-only.
Saving and availability
Changes save automatically. As soon as the agent exists, it appears in:
- The profile dropdown → Agent submenu
- The chat input "+" menu → Agent submenu
- The home screen agent selector
- The Agents page (both Table and Gallery views)
Tips
- Start simple. A short, focused instruction beats a long generic one.
- One purpose per agent. Narrow agents are easier to optimize, share, and delegate to.
- Test by chatting with it. Open a normal conversation with your new agent and verify it follows your guidance.
- Iterate. Most agents get noticeably better after two or three rounds of refinement.
Deleting an agent
You can delete an agent you own from two places:
- The Agents page — row action menu.
- The agent's Properties → Danger Zone at the bottom of the page.
Both flows ask you to type the agent's identifier to confirm. Deletion is permanent and immediately removes the agent from every selector. Existing conversations that ran on it are preserved but lose access to its custom instructions, capabilities, and greeting.
Shared agents can't be deleted from your side — only the owner sees the Danger Zone. To stop using a shared agent, leave the share from Agents → Shared with Me.
Duplicating an agent
You can make your own copy of any agent you own.
There are two ways to duplicate:
- From the Agents page, open the actions menu (⋯) on any agent row and click Duplicate.
- Inside the Agent Builder, open the Properties tab and click Duplicate in the Duplicate Agent section.
A confirmation dialog opens with editable Name and Identifier fields. The defaults append today's date and, if needed, a numeric suffix to keep them unique among your agents. If the source already ends with a date suffix, that date is replaced rather than stacked.
The new agent copies instructions, capabilities, base agent, greeting, diagram and UI preferences. Conversations, versions, triggers, optimizations, sharing assignments, and statistics are not copied — the new agent starts clean. After duplicating, you're taken straight into the Agent Builder for the new agent.
Following the assistant across tabs
The Agent Builder workspace has several tabs (Instructions, Diagram, Properties, Capabilities, Greeting, Triggers, and so on). When you ask the Agent Builder to change something, it normally applies the change in the background and tells you what it did.
If the change you requested is best seen on a different tab than the one you're currently viewing — for example you're on Instructions but the assistant just rewrote the Greeting — it will switch you to that tab automatically and show a small notification, so you can immediately see what changed without hunting for it.
Digital Entities (premium)
A Digital Entity is a premium kind of agent with its own long-term memory, a persistent state, emotions and (soon) the ability to activate itself. Entities cost 2 credits per day.
See the Digital Entities section for the full story: overview and upgrading, Memory, State & Attention and Emotions.