Agents
A fleet that runs the desk.
Nothing leaves without a human.
Globus ships with a set of scheduled agents that read a desk and write up what they found. An admin decides which ones each employee can see. When an employee runs a working agent, the runtime itself forces it into draft mode — the agent writes a brief, not an email.
The rule
It runs on your say-so. It drafts on ours.
The draft-only guarantee is not a setting an employee can flip. It is applied by the code that launches the agent, and the permission is re-checked on the server after the click.
The runtime sets the flag
When an employee runs one of the ten working agents from the dashboard, the launcher writes that agent's own dry-run environment variable to "1" before starting the process. The agent takes the draft path because its environment says so — there is no request field, and no prompt, that can ask for the live path instead. The six reading agents carry no such flag and need none: writing a brief is the only thing they do.
Written to a review queue
A run produces a markdown brief for a human to read. It does not post to a team channel and it does not reach a customer. Scheduled runs of the same agents are a separate path on our side; the button in the portal is always the drafting one. Where that brief currently lands, and why the portal's activity console does not yet show it, is set out plainly in the limits below.
Not trusted from the browser
Authorisation is resolved three times: the admin action validates the slug against the org-scoped set, the run route re-resolves that employee's grants, and the launcher refuses any slug outside the set it was handed. A hand-crafted POST naming a private agent is rejected there.
The assistant cannot start an agent either. The run tool is deliberately absent from the employee chat toolset — an employee fires an agent from the dashboard, not by asking the model to. The model can hand you a query; it cannot hand itself a permission.
What an admin controls
Private by default. Granted one at a time.
An employee signs in and sees nothing. Not a hidden list, not a greyed-out list — an empty one. Agents appear only where an admin has granted them.
Empty until granted
The grant lookup returns an empty set for an employee with no grants, and an empty set on any database error. The failure mode is "you see nothing", never "you see everything".
Only company agents are grantable
Each agent in the registry carries a scope. Only the ones marked as company-scoped can be granted or run in a workspace at all; agents that read the owner's private data are invisible and unrunnable there, and a newly added agent is hidden until it is explicitly marked.
What the console can do
Grant an agent, revoke a grant, set a person's team, set their role, and switch each email agent on or off per mailbox. The console is gated on the admin role and every action re-validates the target server-side.
Two honest notes on granularity. A grant can be aimed at everyone, at a team, or at one person — but on our own workspace today every grant is aimed at everyone, and nobody has a team set, so the team-scoped path is built and not yet in use. And there is no in-product log showing who ran which agent; the guarantee above is enforcement, not an audit trail.
The fleet
Sixteen agents a company can be given today
Six of them read and report. Ten of them do a piece of work and hand it back as a draft. All sixteen are company-scoped, which is the only reason they can appear in a workspace at all.
Reading briefs — they observe and write up, nothing else
- Chief of StaffCross-business daily brief across CRM, product and strategy. It is the business-only twin of the owner's own brief, built without the private inbox and finance sources.
- DronSales desk — pipeline, lead triage, and outreach written as drafts.
- NakulInfrastructure, deploy and scheduled-job watcher.
- VidurPaid-ads spend and performance watcher.
- VyasContent pipeline watcher.
- LomashaCustomer-chat silence watcher — it flags a conversation where a customer is still waiting. It is currently wired to our own EmpMonitor support desk.
Working agents — they produce something you review
- BhishmaMulti-product CRM and sales audit, drafted as a brief.
- ValmikiBlog editorial QA — scores a draft out of 100 and says what has to change.
- VishwakarmaRepurposes an approved blog into LinkedIn and reel drafts.
- SahadevaProposes the next batch of blog topics per product.
- TvashtaTurns an approved blog into a WordPress draft package.
- SamvadaConsolidated daily content-ops digest.
- Competitor Video IntelCompetitor short-form video intel, per product.
- Video ScriptwriterDrafts shoot-ready short-form video scripts.
- Morning Video ScriptsA per-product batch of shoot-ready scripts in one run.
- Video Gen-PromptsTurns a finished script into generative-video shot prompts.
To be exact about the state of our own workspace — the ten working agents are granted; none of the six briefing agents is granted to anyone yet. Granting is the admin's call, not a default we shipped on.
The email agents
Six agents on a shared desk. One of them touches the mailbox.
A shared support address is the one place where a company's work is genuinely queued up in public. Six agents run across those desks — on our own deployment that is eleven mailboxes each, and twelve for the reply drafter — and each is switched on or off per mailbox by an admin.
Labeller Labels only
A fast pass over new mail that files it into a small taxonomy. It labels and it never archives.
Reasoner Flags only
Reads what the labeller left unfiled — the remainder that did not fit a bucket — against the business context, records what it thinks the message is and how urgent it is, and marks it as needing a person. It applies no other label and archives nothing.
Spam rescue Moves mail
This is the one agent that mutates a mailbox. It reads the Spam folder and moves a wrongly-filtered customer message back into the inbox. That is the whole of its power: Spam to Inbox, on a message a classifier judged to be a real customer. It writes an audit row for every move.
Responder Draft only
Where a customer spoke last and is waiting, it writes a reply into that desk's Drafts folder for the person who owns the desk to read, edit and send. Pricing in a drafted reply is taken from a rate card an admin maintains, never invented.
Nudges Draft only
The inverse case: we spoke last and the thread went quiet. It drafts a nudge, chosen to match where the conversation had got to, and leaves it in Drafts.
Learning pass Reads outcomes
Once a night it looks at what the human actually did — which drafts were edited before being sent, which rescued messages were pushed back to Spam — and writes those corrections into a per-desk notes file the agents read back on their next run.
So the honest one-line version: five draft, one moves a message out of Spam. Nothing is sent to a customer by an agent. The drafts sit in the desk's own Drafts folder, in the mailbox the manager already opens every morning, and a human presses send.
What this is not, yet
These are our agents, made visible to your team
Globussoft built this fleet to run Globussoft. That is why it works — and it is also the limit. Read this part before you decide.
Straight answers
- It is not an agent builder. There is no screen where you compose an agent from steps, tools and a schedule. The sixteen above are code we wrote, and several are wired to our own products, trackers and documents.
- Agents for your workflows are a scoped engagement. Bring the process you want automated to a call; we scope it, build it, and grant it into your workspace. That is a project with a person on it, not a toggle.
- No self-serve agent creation. Nothing in the portal creates a new agent.
- No marketplace. There is nowhere to browse or install agents written by anyone else.
- No per-team agent library. Grants can be aimed at a team, but there is no separate library a team curates for itself — one registry, granted out of.
- The activity console does not show briefs yet. A workspace reads briefs from its own directory, deliberately separate from the one our scheduled runs write to, so our internal briefs can never appear in your portal. That separation works. The unfinished half is that a run fired from the portal still writes its brief to the old shared location rather than into the workspace directory the console reads — so the console stays empty and you see the agent cards with nothing behind them. Wiring the two ends together is work we have not done.
- No record of who ran what. Permission is enforced on every run; it is not written to a log you can open afterwards.
If a claim on this page matters to your decision, ask us to show it running. Every mechanism described here is in the code that serves this page, and we would rather walk you through it than have you take it on trust.
Next step
Bring the workflow you want automated to the call
Tell us the desk that eats the most hours — the inbox, the pipeline review, the content queue — and we will tell you honestly whether an agent we already have fits it, whether it is a scoped build, or whether it is not a good use of your money.