Agent Accountability Framework

Questions or comments?

Post them as a comment on the tracking issue -- requires a GitHub account with access to this repo.

AI-DRAFT — earlier-stage than this repo's other two Feature Sets. Seeded from the June 2026 internal concept paper "Agent Accountability Framework: Trusted AI through Human-Accountable Agent Management." Unlike the Policy Advisor and Codex Curator PRDs (v1.5, PM+Lead-Developer-owned, already describing what does/doesn't exist today), this source is an opportunity/strategy paper with an explicit "Not Started" TODO list for nearly every underlying spec (policy violation taxonomy, LangFuse connector, Job Profile schema, productivity baseline). Treat every Feature below as considerably more speculative than the Policy Advisor / Codex Curator Feature Sets — this is Inception-stage framing, not Construction-ready scope. Source doc kept at reference/Novaworks_Agent_Accountability_Framework.docx for durability.

Business case

AI agents are embedded in core HR, finance, procurement, and compliance workflows, but make decisions, take actions, and accumulate organizational impact with no clear ownership, no compliance structure, and no accountability chain. Novaworks — already a Total Work Management platform connected to the HR backbone of the enterprise — is positioned to extend its workforce model to include AI agents: identity, job profiles, human supervisors, and compliance oversight through the same infrastructure that manages human workers.

This is explicitly not an agent-building platform or a technical monitoring system for model performance/latency/hallucination — customers keep using LangChain, Vertex AI, AutoGen, etc. to build and deploy agents. Novaworks's lane is the accountability layer, not the construction layer: managing agents to their expected business outcomes the way it manages human workers.

Three converging forces make this timely, per the source paper: CHROs (91% in the cited 2026 CHRO Survey) rank AI/digitization as their top concern and are becoming AI policymakers, not just technology consumers; regulatory pressure (EU AI Act high-risk obligations enforceable Aug 2, 2026; Colorado SB 24-205; layered CA/NYC/Illinois requirements) is making human-in-the-loop and audit trails legally required artifacts, not just operationally useful ones; and 47% of CHROs report no established AI productivity measurement, leaving the "what did my agents actually do, and were they compliant" question unanswered at board level.

Known gaps, per the source paper's own honest assessment: the policy authoring collaboration model (HRBP + business leader) needs a defined process and tooling spec before it scales; the LangFuse/audit connector is a point-in-time integration that multi-platform agent environments will strain; the minor-vs-major violation taxonomy is underspecified; and the internal audit engine (a Q3 dependency) isn't built yet, leaving the compliance story incomplete without it.

Goal

Give every AI agent operating inside an enterprise customer a workforce- style identity, a job-scoped policy framework, a named human supervisor, an audit trail, and a correction/escalation mechanism when things go wrong — the same way Novaworks already manages human workers. Concretely: an "Agent Worker" identity type in Novaworks, paired with a ServiceNow machine identity, governed by a Job Profile authored collaboratively by business leaders and HRBPs, reporting to a named human, audited against that Job Profile's policy requirements, with minor violations routed to the responsible human and major violations escalated through standard HR paths — up to and including an identity-level kill switch that halts the agent's access pending manual review.

This Feature Set does not cover the separately-scoped Anthropic/Claude partnership thesis in the source paper (reference architecture, shared GTM, workforce-accountability certification) — that's a strategy/BD initiative, not a product Feature, and isn't represented here.