No pre-execution assurance.
Model validation happens before deployment. Audit happens after execution. Nothing checks the individual decision in between, where the consequences start.
Runtime governance for AI agents that move money.
OpenSAFR is an open schema, conformance suite, and reference implementation for SAFR, the industry reference approach for governing agentic AI in financial services at the point of action.
Agents now act. Governance still reviews after the fact.
Financial institutions are moving AI from recommending to executing: payments, trades, credit decisions, filings. The SAFR whitepaper identifies three gaps in how that is governed today.
No human-agent interface.
Human review is often an email alert or a dashboard flag, with no deadline, no defined decision, and no record of what was decided.
Fragmentation.
Where guardrails exist, each deployment built its own. The controls don't interoperate and the records don't share a format.
A shared schema is the fix for the third gap, and it makes the first two auditable. That's what OpenSAFR is building.
SAFR in one picture
Every proposed action is declared, verified, assessed against controls, given exactly one outcome, and recorded.
- Agent proposes action
- Governance Envelopeaction · action trace · context metadata
- Agent Identityverified against the registryfail → Deny + logged
- Controls Repositorywhich controls apply
- Disposition Engineresolves to exactly one outcome
- Auto-Execute→ Financial rails
- Observe→ Financial rails (flagged)
- Escalate→ Human reviewer
- Deny→ Rejected with reason
What OpenSAFR provides
Status: Draft
Governance Envelope schema
A language-agnostic schema for the envelope and audit record, so records from different agents and vendors can be read the same way.
Status: Planned
Conformance suite
Test vectors and a runner that check whether an implementation produces the right disposition and a valid record for a given envelope.
Status: Planned
SDKs
Python and TypeScript libraries for emitting envelopes and validating records.
Status: Prototype
Reference implementation
A working gate with a tiered policy pack, built first for cross-border private banking.
Following the pattern of OpenTelemetry and OAuth: a spec repository, separate SDK repositories, and a public conformance suite.
Built, not just described.
Governed Wealth Intelligence
An agentic wealth-advisory layer governed end to end by SAFR components. The same agent, the same portfolio, and the same market signal across three clients produce three different outcomes: executed, queued for the relationship manager, or blocked. The mandate decides, not the model.
See the implementation →Latest writing
CommentaryGeneral
Introducing OpenSAFR
An open schema, conformance suite, and reference implementation for SAFR. What it is, why the schema comes first, and what it isn't.
Essential reading
BuildFin.ai / MASIndustry
Safeguards for Agentic Finance at Runtime (v1.0)
The source document. Defines the envelope, the four components, and the four outcomes. Start here.
MindForge ConsortiumIndustryCited in SAFR
AI Risk Management: Operationalisation Handbook
Practical guidance including agent-specific practices: least privilege, kill switches, searchable logging. SAFR builds on this groundwork.
IMDARegulatoryCited in SAFR
Model AI Governance Framework for Agentic AI (v1.5)
Singapore's governance framing for agents: bounding use cases, limiting access, human oversight, accountability.
IMFAcademicCited in SAFR
How agentic AI will reshape payments
A three-layer model for agentic payments that identifies control and authorisation as the critical governance point.
Help shape the schema.
SAFR's authors describe it as a starting point that should evolve through implementation experience. OpenSAFR is where that experience can become shared, testable artefacts. We're looking for engineers building agent platforms, risk and compliance teams at financial institutions, and auditors who will have to read these records.