Agent
Payroll agent
Drafts payroll adjustments
Tools
Read
employee_idLook up employee
Read
querySearch policy
New: Introducing the Lemma MCP Server, and what we learned building it.New: Introducing the Lemma MCP Server. Read the blog post
Governance is part of how Lemma runs work, not a layer added after. Each person, team and service gets the actions it needs on each resource, each agent gets only the tools its task requires, credentials and sensitive values stay protected, a person decides when a guardrail trips, and every request traces back to an identity.
Before a Worker or an agent touches a consequential system, someone has to say who can act, what the agent can reach, how secrets stay secret, when a person decides, and how to prove what happened. Here is how Lemma answers each, followed through one payroll exception.
Grant roles on the tenant or on a single resource. Payroll Operations can run the Payroll Adjustment Worker, Engineering can edit it, and its admin can share or delete it. A service account can be granted that one Worker, so an integration runs it without broad administration rights. Every request is checked against those grants, and a refused change is audited.
An agent starts with no tools. Its author adds each one, a function or an MCP tool, with the credential it calls with. A parameter can be pinned to a trusted value from the run: the payroll agent searches policy freely but looks up only the employee on the case, because the Worker supplies that ID and the model never sees the field. The payroll write is not on the list at all.
A Worker names a credential instead of holding it, and stored credential data is encrypted with a key for its tenant. Each API call chooses whose access it uses: one service credential, each user’s own, or, where enabled, one chosen from the run’s data. Mark an input sensitive and protected views of the run mask it, and where enabled, so do supported copies and outputs made from it. Execution still uses the real value, so the calls and files a Worker sends it to remain the author’s responsibility.
Put the check in the Worker, not in the prompt. The agent proposes a payroll adjustment, a rule compares it with policy, and anything over $5,000 goes to a payroll review that waits for approve, correct or decline. The payroll function runs only after an approval, and the same run continues with the reviewer’s answer. You set the condition, who is notified, the paths and what a timeout does.
Security audit events record each covered change: the tenant, the identity that acted and the one it acted as, the request, its status and a trace ID. Refused changes are recorded too. The run’s logs answer the other question: which steps and functions ran. Where enabled, a user permitted to act as a service account can run a Worker as that account, and the audit names both.
The model proposes, the Worker applies the rule, a person decides the exception, and only an authorized function writes. Each part runs under an identity you granted, and every request lands on the record.
See grants, agent tools, a guardrail and the audit trail with our team.
Request a Demo