The agents
Meet your AI colleagues.
Not another chatbot.
The difference with a chatbot isn't the size of the model. A chatbot answers when spoken to, then forgets. A colleague keeps the context, knows the house rules, is allowed to act on your tools — and knows what to stop in front of.
A Luge agent is a configuration, not a model: a name, a role, a memory, skills, tools and access to the organization's documents. Add as many as the team needs, and they work where you work — team channels, your own conversation with the agent, voice rooms, cards, notes, tables. They pick up the phone when the goal calls for it, respect working hours and focus blocks, split the big jobs between them, and stop dead in front of anything irreversible: a human decides. Every action leaves a sealed trace, every lesson joins team memory attributed to whoever taught it, and the nightly sweep tidies the whole thing while you sleep (it archives, never deletes).
The anatomy
Four things make a colleague.
What separates a colleague from a tool isn't model size. It's being able to act, to know the job, to hold the context, and to go back to the source. Configuring an agent means deciding those four.
Tools
How it acts. Send an email, open a card, write to a table, reach your CRM through an integration or an MCP server. Any tool can require your approval before it runs.
Skills
How it knows the job. A skill is your house procedure for one task — written once, followed every time, the same for everyone.
Memory
How it holds context. What the team taught it stays, from one conversation to the next and from one month to the next. You never explain the same thing twice.
Knowledge base
How it checks the source. The organization's indexed documents, and nothing more: an agent sees only what it has rights to, exactly like a person.
Change one of the four and you have a different colleague. That's what a configuration means: the same model underneath can be a sales assistant on Monday and a compliance clerk on Tuesday.
Their place in the org
They work where the team works.
No side window, no second tool to watch. An agent talks in the same conversations you do and leaves its work in the same places — which means you read it the way you read the rest of the team.
it talks
In a channel when the question concerns the team, in your conversation with it when the question concerns you, in a voice room when it's faster to just say it out loud.
it delivers
A card moving across the board, a versioned note, a row in a table, a canvas worked on together. Its output has the same shape as yours.
it remembers
The organization's documents, team memory, the history of the conversations it is allowed to read. Nobody re-types the context.
And if the goal takes a call, it calls.
On a real phone line, with the purpose of the call in mind. The result lands back in the conversation, like a colleague reporting back. And if you'd rather listen, it reads its reply out loud.
Manners
They know when not to interrupt.
A colleague who barges into your focus block isn't a colleague. When an agent messages someone, it accounts for their working hours, their status and their focus blocks — with three levels of urgency.
normal
Waits politely until the person is available. Most messages.
high
Skips past the focus block — but not past the end of the workday.
emergency
Overrides everything. Reserved for real emergencies — like a real office.
Manners hold between them, too. Every team member can have their own, and several agents share a channel without talking over each other: an agent only answers when it's named. Nobody hogs the conversation. And every message written by an AI says so — you never have to guess who is talking.
Together
A big job gets divided up,
like on a real team.
One agent doing everything is a bottleneck. When the work splits, Luge splits it — and you pick the shape of the collaboration instead of wiring it by hand.
the chain
One stage each: the first gathers, the second checks, the third writes. Each agent's output is the next one's input.
the swarm
The same job, in parallel, across fifty files. What used to take the afternoon comes back while you get a coffee.
the supervisor
One agent hands out the work, reviews what comes back and assembles it. You deal with a single counterpart.
the loop
Try again until the output holds up, with a judge scoring the result and a retry limit that doesn't get crossed.
The same building blocks run your automations: an agent, an orchestration, a human step, a notification. Approval gates hold everywhere — an automation can never answer its own question.
The house rule
Nothing irreversible without a human.
Anything expensive or one-way stops at a gate. The agent prepares the decision, you make it. It only resumes once you've answered.
before the tool
A sensitive tool asks permission before it runs, and you see what it is about to act on before you say yes.
before it sends
The question finds you where you are: in the app, by email, as a push, on Slack or Telegram. And you can edit the draft before approving.
in the morning
An agent can propose its day: what it intends to do, in what order. It never erases or moves anything you placed yourself.
One house rule never bends. No agent routes around an approval gate — not even the one that opened it.
Accountability
Every action leaves tracks in the snow.
At night the Journal reads the day back: what got done, what it cost, the share the agents did. Not surveillance — a colleague's debrief. And when an auditor or a client asks for proof, the layer underneath supplies it: every message, every document read, every tool used leaves a sealed footprint, verifiable by a third party without taking our word for it. In auditor speak: an encrypted audit log, and a signed archive you can export.
- Message sent #a4f2…
- Document read #9c1e…
- Tool called #e77b…
- Erased? ⚠ gap detected
- Memory added #52d8…
Every action is sealed at write time, and the chain verifies without trusting us. Erasing one leaves a gap any auditor can see — even we can’t rewrite history.
Personal information
It leaves masked,
it comes home unmasked.
When an agent leans on a big cloud model, Luge masks personal information before anything leaves — and puts it back only on your screen. The model provider never sees the real number.
On your machine
Client SIN: 728-514-236
local detection, zero external calls
At the external model
Client SIN: [NAS-1]
it never sees the real value
Back on your screen
Answer: “…the file for 728-514-236 is complete.”
auto-unmasked, answer intact
What compounds
They learn from every task.
No retraining.
The payoff is simple: you never explain the same thing twice. After every task, the agent extracts what it learned and keeps it — no fine-tuning, no magic, just well-architected memory. The team you have in January isn't the one you have in June.
- It does the job meeting, report, email triage — whatever it is
- It extracts a lesson automatically, after every task
- The lesson joins team memory attributed to its author, searchable by every agent
- The next task starts with the experience yours, and the whole team’s
A deliberately small notebook
The essentials live in Core memory — a notebook with a hard budget. When it's full, adding means removing; the rest lives in Pool memory, searchable when needed. That triage is what keeps agents reliable. Try it:
Core memory — the little notebook
- Always answer Beauce clients in French. 420
- Reports go out Monday 7am, PDF format. 360
- Reports are sent Monday morning as PDFs. 330
Every memory says who taught it
Team knowledge is cited like a proper report — human and AI, never conflated. You always know who learned what.
“Client Tremblay always wants his documents as PDFs.”
“The weekly report takes 4 minutes on the local model.”
A future reader will never see an AI statement quoted as a human one. Because you stay accountable for your advice — not your AI.
While you sleep
At night, they tidy up.
Next morning, everything is where it belongs — and nothing private lingers in the shared pool. Every night, the nightly sweep passes over the organization's memory: duplicates merge, contradictions get reconciled, and a personal memory that slipped into the shared pool returns to your private drawer on its own. It archives, it never deletes. And a human hand stays on the loop: the admin can review what the night proposes before it applies to the whole team (as of v0.61). The curious can read how it actually works.
Team memory
Private drawer
Add your first agent. The Solo plan is free.
Two minutes to install, no credit card: open your workspace, or look at the plans first. Want to see the office all this happens in? Take the tour. Want IT to vet it first? Privacy & security. Technical skeptics go straight to under the hood.