Governance & audit
Autonomy you can defend. Every action an agent takes carries its authority, its evidence and its author, on an immutable record.
Audit trail
append-only · queryable · exportable
TIME
AUTHOR
ACTION
AUTHORITY
EVIDENCE
14:02
Atlas agent
Mission created
Workflow: Tendering v3
RFP-2216
14:07
Atlas agent
412 pages parsed into 247 tasks
Read: document store
412 documents
14:18
Atlas agent
219 responses drafted from company evidence
Read: approved sources
638 citations
14:32
Sara Al-Otaibi
Payment terms approved at 75 days
Role: commercial director
Policy 3.2
14:33
Atlas agent
Submission assembled, ERP updated
Write: ERP, scoped
Audit ref 8841
the record is written as the work happens, not reconstructed after it
Every action carries three things
Agent or human, read or write
Its authority
The permission, role or workflow rule that allowed it to happen at all.
Its evidence
The sources, citations and inputs the action was grounded in.
Its author
Who or what performed it: a named person or a named agent, never anonymous.
The governance plane.
Five controls apply to every mission, every retrieval and every model call. None of them are optional.
Permissions
People and agents act only within the access you grant.
Masking
Sensitive fields are hidden before any model sees them.
Immutable audit log
Append-only record of every action, decision and access.
Spend controls
Budgets and limits per workflow, team and department.
Observability
What ran, where, at what cost and with what outcome, live.
Human controls
Autonomy is a dial, not a switch. You decide which actions run free, which need review and which always stop for a person. The boundary is enforced by the platform, not by convention.
Governance is one part of the layer.
Foundry pairs it with model routing, grounded retrieval and context & memory.