How it works
How the agents do the work, and where your team steps in
Belowdecks is an agentic operating system: software that puts AI agents to work on your company’s tasks. Something puts the agent to work, it does the task, and your team steps in to answer its questions and approve the actions chosen with you.
Who does what
Who acts, and where
Your customers and suppliers write in from outside, the agents work in Belowdecks, your team decides, and our team installs and monitors the whole thing.
- Step 1 of 4: Marc fills in your website form. Who acts: Marc, Customer
- Step 2 of 4: The agents prepare the quote. Who acts: Agents
- Step 3 of 4: Julie approves from an email. Who acts: Julie, sales, Your team
- Step 4 of 4: The quote goes out and the request is recorded. Who acts: Agents
Marc, Customer
Outside,Writes to your company
He sends an email or fills in your website form, without ever seeing your Belowdecks workspace.
- Hello, I need a quote for three buildings.
- Quote form sent from your website
Julie, sales, Your team
In your company,Decides
Julie answers the agents’ questions and approves important actions, in Belowdecks or from an email.
- Answers from an email
- Approves in Belowdecks
Agents
In Belowdecks,Works
Each agent works in its own space, with its tools and a time limit, then finishes with a result.
- Front-desk agent
- Quoting agent
Belowdecks team
Behind the scenes,Sets up and monitors
We install your workspace, set up the agents with you and monitor their work. Decisions stay with your team.
- Installs your workspace
- Keeps each agent’s space separate
- Monitors
In four stages
How a task unfolds
Every task handed to an agent goes through these four stages.
- 1
Something puts the agent to work
An email, a file, a signal from one of your software tools or a scheduled time. Someone on your team can also start the agent themselves.
- 2
The agent does the work
It follows its instructions, written in plain language, and relies on the AI of a provider such as Anthropic or OpenAI. Your team can follow each step live.
- 3
Your team answers and approves
For the actions chosen with you at installation, the agent asks someone on your team for their go-ahead before acting. If information is missing, it stops and asks its question.
- 4
The result is recorded
Every task leaves a log and ends with one of three results: Delivered, Partial or Blocked. An activity feed shows what is in progress, what failed and what is waiting for a person.
What starts the work
Six ways to put an agent to work
Each one starts an agent, or an automatic task when no judgment is needed. An automatic task is a small program that redoes the work without AI.
-
An incoming email
Each new message in a mailbox, filtered by sender or subject if needed.
-
A news source
Each new article from a source you follow: trade press, a public agency, alerts.
-
A file dropped in a folder
A file added to or changed in a folder.
-
One of your software tools
With Passerelle and Flotte, a tool reports an event, such as a new order, through the API or a webhook. Belowdecks checks where it came from and filters out what doesn’t concern you.
-
A schedule
Described in plain words, such as “weekdays at 8 a.m.” or “the 1st of the month”.
-
A new row in a list
A request added to a list, by an agent, a function or your team, can put the quoting agent to work.
The agent’s space
What the agent has to work with, and its limits
Each agent is set up once with our team, then works on its own within these limits.
Tools
It reads and writes files, runs small programs, browses the web and uses your software when the software allows it.
Instructions in plain language
Your ways of working, written once: the tone of your replies, your filing rules, your checks.
A time limit
Adjustable for each agent, it stops any task that reaches it.
A separate space
Each agent sees only its own files. Our team switches this separation on and checks it during setup; without it, no task starts.
Encrypted passwords
Passwords and access keys are stored encrypted, masked in logs and never exported.
Agents that share the work
An agent can start another one or split a task. The agents it prepares stay switched off until your team turns them on, unless you have given it that autonomy.
Go-aheads and questions
The actions you choose go through your team
In a conversation, the agent first presents an approval request: a pre-filled form you can read in seconds. When an agent starts on its own, for example when an email arrives, we can also set it up to wait for your go-ahead before acting.
What goes out is what you read
The request carries the exact values the action will use. Depending on the action, you can correct a field before approving; a refusal cancels the action.
Who approves what
Someone on your team, based on their group’s permissions. You choose, agent by agent, who receives the approval requests.
An answer from an email
A link received by email lets your team approve or answer without signing in, after a confirmation page.
The agents’ questions
An agent missing information stops, asks its question and resumes once your team answers. The other agents keep working in the meantime.
Emails to your customers and suppliers
We set up how they are sent with you during installation. The agent’s instructions can have it ask for your go-ahead before each email.
A trial before going live
Each agent first handles real cases with the instruction to send or change nothing, in a workspace deleted at the end. Your team reads the report of what it would have done.
Send Atelier Lavoie a reminder for invoice 2041?
- Customer
- Atelier Lavoie
- Invoice
- 2041, overdue since September 30
- Tone
- Friendly, no late fee
Choices offered: Edit, Approve.
Over time
How Belowdecks improves from one week to the next
Every task leaves a trace, used to fix what fails and to hand repetitive work to automatic tasks.
-
An analysis every night
Belowdecks spots repeated failures, time limits reached, unusual AI usage and go-ahead requests left unanswered. An AI analysis also looks for the cause of a problem, agent by agent, and suggests a fix.
-
A review every Monday
The week’s findings arrive in your conversation with the captain, the assistant you use to run your agents. A precise fix, such as a longer time limit, is suggested to you, and you accept it or not.
-
Automatic tasks for repetitive work
When an agent repeats the same steps, the analysis flags it. An automatic task can then be written, often with our team, and tried on examples. Your team approves putting it into service, and Belowdecks refuses it if a trial fails. If the task fails later, the agent takes the work back.
-
An agent that suggests changing its instructions
An agent can suggest changes to its own instructions, which your team accepts or rejects. The captain can also make a change at your request, after showing you the before and after.
-
A report on the 1st of the month
Administrators receive the month’s summary, agent by agent: work done, results, AI usage, questions and pending findings.
Example
The Monday review, step by step
Julie reads the weekly review in her conversation with the captain, accepts a fix, then puts into service an automatic task the Belowdecks team prepared.
Who acts in this example
- Customer Writes to your company
- Your team Decides
- Agents Work
- Belowdecks team Sets up and monitors
Julie accepts a fix on Monday, then puts an automatic task into service
Belowdecks workspace
Julie, sales, Your team Your teamJulie receives the agents’ questions and what needs approval here.
The agents
Agents Agents Now-
Monday, 8 a.m. The weekly review arrives in Julie’s conversation with the captain
Step 1 of 7 Monday, 8 a.m. Belowdecks workspace
Julie and the company Entretien Boréal are fictional.
Ways in
Who uses Belowdecks, and how
The same agents and lists serve everyone. Permissions depend on who is talking to Belowdecks.
-
The team space
Your team talks with the captain, follows work in progress and answers the agents. With Passerelle and Flotte, screens are added: lists, action buttons, figures and charts.
-
Permissions by group
Each group on your team can hold conversations, read or edit data, follow the work, start an agent or answer the agents.
-
Email
Your team answers the agents’ questions and approves their work from an email, without signing in.
-
Your software and AI assistants
With Passerelle and Flotte, they can start an agent, follow its work, answer a question and read or write your lists. The technical details are on the Developers page.
Questions about how it works
Who on our team answers the agents and approves their work?
One or more people you choose. You decide, agent by agent, who receives the questions and who can approve, based on each group’s permissions.
Can the agents act in my software?
Yes, when your software allows it, in every plan. Our team sets up the connection during installation, often through an MCP server, and stores the access keys in the encrypted vault. We choose with you the actions in your software that wait for the go-ahead of someone on your team.
For your software to launch an agent itself, through the API or a webhook, you need the Passerelle plan.
What language does Belowdecks work in?
The agents write in the language set by their instructions, French or English. They answer your customers by email in the language the customers use, if their instructions say so.
Next step
Let’s see what work to hand to agents at your company
A free 20-minute first call is enough to know whether Belowdecks is a fit. At the diagnostic, we then pick a first task with you and the person who will answer the agents, and you receive a written plan.