Capability layer: AI agents
Organisations hand real work only to agents with a short list of what they may touch.
Inside those limits the agent carries out the steps itself, and every action it takes is recorded. The limits belong to the organisation: which tools and systems each agent may reach, and where it waits for a person to approve.
What happens between a request and a result
An agent that reaches for a tool before it understands the task is guessing with real systems. So the order below is fixed, the approval step included.
- 1
Reads the task
Before anything moves, the agent works out what is being asked, of which record, and why. A missing detail becomes a question, not a guess.
- 2
Reaches only permitted tools
It cannot reach a system nobody gave it. The tools an agent may use, and whether it may read or write through each, are fixed when it is deployed.
- 3
Takes the step
Rather than suggesting the step, it takes it: the record changes, or the document is prepared, in the system your team already uses.
- 4
Stops for approval
It stops and shows a person what it intends to do. Some steps never run on the agent’s judgement alone, and the organisation decides which.
- 5
Hands over at its limit
Past its limits, it does not improvise. The case goes to the right person with everything that happened so far attached.
What is configured around each agent
Agreed at scoping, deployment by deployment: which of these the agents use, and how each one is set.
Tools and permissions, set per agent
Whether an agent may only read a record or also change it is decided tool by tool, and it gets no tool it was not assigned.
Memory, triggers and retries
A failed step is retried within set limits before anyone is alerted. Between steps the agent keeps the task’s context, and it can start on an event, such as a new request, or on a schedule.
A record of every action
Operators watch agents while they run, and review afterwards which tool was called, what changed and which task it served.
Several agents, one process
Each agent keeps its own role and its own tools, and the work passes between them in a defined order when one process is too long for a single agent.
Workspaces and roles
Staff sign in with their organisation accounts through single sign-on, and each person’s role decides what they reach inside their team’s workspace.
Where agents do the work
Agents your customers talk to
The agent holds the customer conversation until someone on your team is needed, then hands it over. That is ChatSheba’s job, with a page of its own.
Agents for work inside the company
An agent can take on work no customer ever sees, within the tools assigned to it: gathering what an internal request needs, or preparing a file before someone reviews it.
Agents that move work between steps
Every approval point is a stop. Between those stops, the agent carries a process from one system and team to the next.
Where an agent stops
Changing those limits is a person’s decision, and the change is recorded. The agent itself works inside defined tools, permissions and escalation rules, and it cannot widen them.
FAQ
Before an agent works in your systems
Begin with the process your team repeats most.
We work out with you which steps an agent takes and which stay with your team, starting from how the process runs today, stops included.
Tell us the process to turn into an agent