Agents that do the work, not just the answering
Describe a job in plain language. Mantaq builds an agent that reads your company's knowledge, acts in the tools you have connected, and hands back something finished. You see every step it took.
Four things make an agent
You write the first one. Mantaq proposes the rest, and you change anything you disagree with.
Review every inbound vendor contract against our clause checklist, flag anything outside policy and post a summary for the legal team.
What happens when an agent works
Every run happens in a sandbox that is created for that run and thrown away afterwards. Nothing an agent does leaks into the next job.
It gets a clean sandbox
An isolated environment is started for the run and destroyed when it ends. The agent can install what it needs and write files without touching anything permanent.
It never sees your credentials
Outbound calls go through an egress proxy that attaches the right credential on the way out. The agent cannot read tokens or secrets, and traffic that tries to bypass the proxy is dropped rather than sent.
It has to cite its sources
Agents are instructed not to state anything that is not grounded in something they retrieved. When the data is not there, the agent is required to say so rather than fill the gap.
It can act, not only read
An agent can post to a channel, draft an email, update a record or produce a document or a working web app as its deliverable. Reading is the starting point, not the limit.
You see the working
Each run shows the steps taken and the sources behind them, so a result can be checked rather than trusted. That is the same reasoning trail you get in ordinary chat.
An agent asks before it acts
Reading is one thing. Sending an email or changing a record is another, so those actions are gated until somebody says yes.
Approval before the action
When a run reaches an action that leaves Mantaq, it pauses and notifies someone who can approve it.
Or approved in advance
Routine targets can be approved once so the agent stops asking about them. Anything outside that list still stops and waits.
Runs on a schedule
Give an agent a schedule and it keeps the job current. Every run is recorded, and a failure notifies the owner rather than passing quietly.
An action you decline is not retried. The agent reports the outcome and offers an alternative.
Secrets are attached by the proxy on the way out, never handed to the agent.
Runs, their steps and their outcomes are kept, so you can see what an agent did and when.
Built by one, used by the team
One person works out how a job should be done, and everybody else gets the result. Share an agent with a person, a group, or the whole workspace.
Access comes in two levels on top of ownership. Editors can change how an agent works. Viewers can run it and inspect its setup, but not alter it.
Start with the job nobody wants
Pick the task your team does every week and resents every week. That is usually the one worth handing over first.