Set once, enforced every time
Rules are set by your team and enforced by the system on every run, not left to the prompt.
Governance and audit
Control built into every run, not added later.
Decide who each agent is, what it can touch and when it must stop. Every action is logged, so you can always see what happened.
Comparing approaches? Read governed vs ungoverned agents.
Controls RBACSkill accessModel policyApproval rules
Audit log
User approved packet
Agent called CRM tool
Knowledge source checked
Six controls, on for every agent.
People can step in at any point
Any run can be paused or cancelled, and hard cases go to a person instead of a guess.
Nothing is a black box
Every decision and action can be traced back, from the sources used to the person who signed off.
Identity and access for every agent
Every agent acts under a name, an owner and a scope. It can only use the connectors and data you approve, and nothing is open by default. Role based access decides which people can build, run and review agents.
- A named owner for every agent
- Every connector is governed
- Role based access for people
Policy and model rules
Guardrails are set at the organization level and enforced before work runs. Decide which models each team can use, which actions an agent may take on its own and which must wait for approval.
Low risk draft Fast model
Legal reasoning Approved advanced model
Sensitive data Private provider
Central policy
Controls cost, quality, latency, privacy, and task complexity.
Supervision and escalation
Any run can be paused or cancelled at any time, not only at the start. When an agent reaches a step it is not allowed to take alone, it stops and hands the decision to the right person. If an approval waits too long, the agent sends reminders and can alert an admin.
- Pause or cancel any run
- Hard cases go to the right person
- Reminders when approvals wait
- Read the refund request
- Check the order history
- Refund above team limit
Escalated Needs a decision from Priya, Support lead
The agent paused this step. Reminder sent after 2 hours.
Approve Pause run Cancel
Audit log and output history
Every decision and action can be rebuilt after the fact: the request, the sources, the model, the tool calls and each human sign off. The record sits next to the work and is ready for review or audit.
- 09:41 Request classified
- 09:43 Scope checked
- 09:48 Evidence gathered
- 10:02 Approval waiting
Artifacts Contract delta.pdf Spend summary.csv Recommendation.md
How it works
From request to recorded result.
-
Name the agent and its owner
Every agent gets an identity and a person responsible for it.
-
Scope its access
Choose the connectors and data it may use. Nothing else is open.
-
Set policy and approvals
Decide what it may do alone and what waits for a person.
-
Supervise live runs
Watch, pause or cancel work as it happens.
-
Review the record
Every run leaves a full trail for review and audit.
FAQ
Governance and audit: questions and answers
Short answers to what teams ask before they put this part of SoftworkerAI to work.
Contact us directlyWhat does governance cover in SoftworkerAI?
Six controls: identity, access, approvals, supervision, audit and policy. They apply to every agent and every run, and governance is included on every plan, never sold as an add on.
Are the rules enforced, or only suggested to the model?
Enforced. Policies are applied by the system at runtime, not written into a prompt. An agent cannot use a tool, read data or take an action that its policy does not allow.
Can we stop an agent in the middle of a run?
Yes. Any run can be paused or cancelled at any time by the people allowed to supervise it.
What happens when an agent is not allowed to do something?
It escalates. The agent stops at that step and asks the right person to decide, with the context attached, and the decision joins the record.
Can we use our own model provider?
Yes. You can bring your own LLM provider and set model policies for each team or workflow.
The multiplayer workspace for humans and AI agents is how teams move beyond prompting and safely hand AI real work. It gives people one place to plan, approve, and oversee execution, so they can run work with AI agents they can govern and trust.