Creating an agent
The four creation paths, choosing the AI model, writing a good mission, and the dry run before the first real execution.
7 min de leituraAtualizado a 7 de agosto de 2026
Esta página ainda não está traduzida para o seu idioma: é apresentada em inglês.
Four paths lead to the same result. Pick according to how much you already know about your need.
The four paths
| Path | When to pick it | Time |
|---|---|---|
| Start from a template | Your need resembles a common case | 2 minutes |
| Describe it in plain language | You know what you want, not how to set it up | 5 minutes |
| Configure it by hand | You want to control every rule | 15 minutes |
| Import a file | You are bringing over an agent exported from another workspace | 1 minute |
Start from a template
The gallery offers eight proven agents, with permissions already trimmed to the strict minimum. You pick one, adjust the mission and the schedule, and it is ready.
This is the recommended path for a first agent, even when the template does not match exactly: adapting a well-framed agent is easier than setting everything up yourself.
Describe it in plain language
You write what the agent must do, and Aidoo proposes a full configuration: rewritten mission, AI model, trigger, rules, limits.
Every Monday morning, find quotations above €10,000 sent more than two weeks ago with no answer, and schedule a follow-up activity for the salesperson in charge.
The proposal always goes through your review
Aidoo never creates the agent directly. You see the proposed configuration, you change it, then you approve. This generation consumes a few credits, like any model call.
Configure it by hand
The full form, section by section. Useful when your case is specific, or when you want to understand exactly what each setting does.
Import a file
An agent exports to the aidoo-agent/v1 format and imports elsewhere. Handy for reusing a proven agent in another workspace. The export carries no Odoo identity and no secret: you designate the member again on import.
Writing a good mission
This is the setting that matters most. A vague mission produces an unpredictable agent.
Say precisely what to look for. "Overdue invoices" leaves too much room; "validated customer invoices, unpaid, past their due date" leaves none.
Say what to do with each item. The agent needs to know whether it adds a note, schedules an activity, sends an email, or simply counts.
Say how to conclude. "Finish with a summary: number of overdue invoices, total amount, top five customers concerned." Without that sentence, the reporting varies from one run to the next.
Bound the volume. "Process at most the 20 oldest" prevents a run from taking on 400 records the day your database grows.
The read-aloud test
Have someone unfamiliar with the need read the mission. If they ask you a question, the agent will hit the same question, and it will answer it on its own, its own way.
Choosing the AI model
Each model shows three markers: cost, speed, reasoning ability. Cost translates directly into credits consumed.
| Type of mission | Suggested model |
|---|---|
| Listing, counting, summarising a list | An economical model |
| Writing a message adapted to context | A mid-tier model |
| Analysing, comparing, choosing between options | The most capable model |
Start with an economical model. If results are rough, move up one notch: that beats paying for a powerful model to count invoices.
During the free trial
Only economical models are available, and the trial is capped at 500 credits or 30 days, whichever comes first.
The trigger
Schedule: a simple time expression, or a guided choice ("every working day at 8 a.m."). No tier bounds that frequency: the agent follows exactly the expression you give it, in its own timezone.
Event: you pick an Odoo model, the event type (creation or update), and optional conditions. Aidoo watches that model at a rhythm set by your tier: one hour on Starter, a quarter of an hour on Pro, five minutes on Business. That is the agent's maximum reaction delay.
For instance: model crm.lead, on creation, with no condition, to qualify every new lead. Or model sale.order, on update, with the condition state = sale, to react to confirmed orders.
The Odoo identity
You designate the Aidoo member whose Odoo rights the agent will borrow. Only owners and admins can make this choice, and only members who have saved their own Odoo credentials appear in the list.
No extra Odoo licence needed
An agent runs on any existing Odoo licence: it borrows the account of a member already in place, with no extra user to create and no licence to pay for. That account sets its scope, so pick it according to what the agent must be able to touch.
The safety limits
Four guardrails, adjustable per agent:
| Limit | Purpose |
|---|---|
| Tool calls per run | Stops a loop from consuming endlessly |
| Tokens per run | Caps the cost of a single run |
| Maximum duration | Interrupts a run that drags on |
| Writes per run | Blocks the agent beyond N changes |
The write cap is the most useful early on: set to 5, it lets you see what the agent changes before you give it free rein.
You can also set a monthly credit budget specific to the agent, on top of the workspace pool.
Test before activating
Always do a dry run
Test mode runs the mission for real, but simulates writes: the agent shows you what it would have created or changed, without touching Odoo. It is the only way to see how it interprets your mission before it has any effect.
Every run drops its report into the agent's discussion thread: what it observed, what it actually wrote, and the figures that sum up its pass. That is the first thing to read after a test.
For the detail, the debug icon on a run bubble opens the full trace: every tool call, with its parameters, result, status and duration. If the agent reads the same record ten times or wanders onto an unexpected model, the mission needs rewording.
Activate
Once the test is convincing, activate the agent. It then mints its own API key, whose permissions derive exactly from its rules: no more, no less. That key is revoked automatically when the agent is archived.
The first real run deserves a careful read. After that, the daily digest is enough.
Next step
- Rules and permissions: bounding what the agent may touch
- Credits and consumption: measuring and controlling cost
- Best practices and troubleshooting: when an agent misbehaves