Repository navigation
Design question: persistent organizational decision context above stateless ACS controls #4200
Replies: 4 comments 1 reply
|
This separation looks consistent with the current ACS model. The security model explicitly says the runtime retains no mutable state that influences future verdicts, and that each evaluation operates on the snapshot supplied by the host. So I would keep the long-lived organizational state outside ACS and project only the context needed for that intervention into the snapshot/annotations. For provenance, I wouldn't copy the whole organizational record into ACS. I'd carry stable references or content hashes with the projected context, then correlate the resulting decision back to the external record. There are already a few useful pieces for that in AGT: evidence can carry a content-addressed That gives roughly:
Whether that deserves an official reference integration is probably a maintainer call, but the boundary itself seems to fit the current stateless ACS design rather than requiring ACS to own persistent organizational memory. |
|
Dear Imron,
Thank you for your thoughtful and helpful response. Your clarification
on keeping long-lived organizational state outside ACS, while
projecting only the context required for each intervention, helped
clarify the architectural boundary we were examining.
To make the remaining boundary more concrete, would the following
host-side pattern be consistent with the intended AGT/ACS model?
Long-lived organizational context
→ Versioned context projection / projection manifest
→ Stable references or content hashes + provenance pointers
→ ACS snapshot / annotations
→ ACS decision / evidence
→ decision_id / trace correlation
→ Organizational audit / memory update
The aim is to make the exact organizational context used for a given
evaluation reconstructable later, without copying the full
organizational record into ACS.
Is there already a preferred AGT extension point or annotation
convention for carrying a stable projection/version identifier for
that context? Or would you recommend keeping the projection manifest
entirely external to ACS/AGT and correlating it through the existing
evidence/audit identifiers?
I’m trying to preserve ACS statelessness while keeping the
organizational decision context reproducible and auditable.
Thank you again for your guidance.
Best regards,
Ali Anvari
Founder, METAS
Reference: DCT-REF-K1-MICROSOFT-ACS-DISCUSSION-4200-HANDOFF-20261003-001
Case: GitHub Discussion #4200 — microsoft/agent-governance-toolkit
Owner Runtime: K1 | Business & Ecosystem Development
|
|
Dear Imron,
Thank you — this resolves the boundary very clearly.
The distinction between keeping the versioned projection manifest
external while carrying only a stable projection reference/hash in
host-provided snapshot metadata is particularly helpful. The
reconstruction chain you outlined also gives us a clean way to
preserve statelessness without losing organizational auditability.
I appreciate the clarity and the time you took to work through this with me.
Best regards,
Ali Anvari
Founder, METAS
…On Sat, 03 Oct 2026 05:27:09 -0700, Imron Reviady ***@***.***> wrote:
Yes, that host-side pattern is consistent with how I read the current contract.
One distinction I'd make is that I wouldn't use ACS annotations as a generic place for the projection identifier. In the current snapshot contract, annotations has a specific role: it contains outputs from annotators declared in the manifest.
I'd keep the full projection manifest external, but if the exact projection is part of what made the decision reproducible, I'd put a stable reference and/or content hash into the explicit snapshot, for example under host-provided metadata:
{
"metadata": {
"org_context": {
"projection_id": "ctxproj-123",
"projection_hash": "sha256:..."
}
}
}
The external store can then resolve that identifier to the full versioned manifest.
For the output side, if that projection is itself evidence, evidence.artefact / verification_pointers look like reasonable places to reference it. Otherwise I'd keep the projection reference in the organizational audit record and correlate it with the ACS result through decision_id and the trace/session identifiers.
I don't see a standardized projection_id field in ACS today, so I would treat the naming/schema above as host-owned rather than an AGT convention.
That gives the reconstruction chain:
projection id/hash -> external manifest -> explicit ACS snapshot -> decision/evidence -> decision/trace correlation -> organizational audit
This keeps ACS stateless while still making the exact decision context reproducible.
—
Reply to this email directly, [view it on GitHub](#4200?email_source=notifications&email_token=CQGDK2AAN6RLHDGA47XTZQ35SDWB3A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBXGMYDKNZXUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-18730577), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/CQGDK2EU7CTTCELAJMIOICL5SDWB3AVCNFSNUABJKJSXA33TNF2G64TZHMYTCNZRGEYTCNBWG45UI2LTMN2XG43JN5XDWMJQHEZDKMJZG2QXMAQ).
Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for [iOS](https://github.com/notifications/mobile/ios/CQGDK2H77OZZR2HTNVPF77D5SDWB3A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBXGMYDKNZXUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVJTG633UMVZF62LPOM) and [Android](https://github.com/notifications/mobile/android/CQGDK2HPXSZ24DW2UHKUFVT5SDWB3A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBXGMYDKNZXUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVZTG633UMVZF6YLOMRZG62LE). Download it today!
You are receiving this because you authored the thread.
|
|
I agree with the existing distinction: authority should be portable independently of the model, but the replacement runtime must not inherit permission merely because the previous runtime possessed it. A useful way to make that boundary testable is to define an explicit authority-continuity contract for each intervention point. Suggested host-side contractFor a proposed action, the host assembles:
The receiver independently validates the authority against the exact action before committing an effect. A runtime migration should change runtime provenance without implicitly changing the underlying mandate. But an unchanged mandate ID must not imply that its authority is still valid. Critical testsI would use matched scenarios with the same proposed action:
These tests should capture both the policy outcome and the observed receiving-system state. Evidence boundaryI would distinguish:
For replay and historical verification, retain the versioned mandate, relevant revocation evidence, runtime identifiers, exact action digest, and decision correlation identifiers. Recommendation: ACS need not own the portable mandate lifecycle. The host can project the current authority result into each stateless intervention snapshot, while the receiving enforcement boundary independently checks authority freshness and action binding. The useful reference integration would be a minimal two-provider migration test with a revocation race, a parameter-mutation contrast, and an independently verifiable execution result. That would demonstrate portable authority continuity without conflating provider identity, previous authorization, and actual execution. |
Uh oh!
There was an error while loading. Please reload this page.
ACS intentionally keeps policy evaluation stateless, with the host supplying a complete AgentControlSnapshot at each intervention point. I am exploring an organizational governance pattern in which longer-lived context—mission ownership, prior decisions, evidence provenance, authorization state, human approvals, and cross-case learning—is maintained outside the agent-control layer and selectively assembled for each policy evaluation.
The intended separation would be:
Persistent organizational decision context
→ host-composed ACS snapshot
→ ACS policy decision
→ agent/tool execution
→ execution evidence
→ organizational audit/memory update
I would appreciate guidance on three design questions:
Is this separation consistent with the intended ACS architecture, with persistent organizational context treated entirely as a host concern?
Are there recommended patterns for linking the provenance of externally assembled organizational context to an ACS evaluation and its resulting evidence/audit trail?
Would a reference integration pattern for this boundary be useful within the Agent Governance Toolkit documentation, or is it intentionally outside the toolkit’s scope?
This is an integration-architecture question rather than a request for ACS to become stateful.
All reactions