Most agent governance watches tool calls: what the model requested, what the gateway allowed, what the workflow executed. Four November 2026 preprints suggest that lens is incomplete. Long-running agents also move state: they compact context, consolidate memory, roll back workspaces and resume after a pause for approval. Those state transfers are where approvals can be lost and where tool-call logs see little.
The State-Transfer Blind Spot
A preprint submitted on November 17 argues that "while conventional execution traces log model prompts and tool calls, critical security and reliability failures occur during state transfers" (Shah and Kathrotia, arXiv:2611.23346). It names three failure modes: a transfer can admit unchecked content, rewrite state lossily (for example, compaction dropping user constraints), or restore state incompletely (a workspace rollback that leaves aborted state active). The paper proposes an audit point at each transfer boundary that emits a hash-chained record. Cheap deterministic checks, such as comparing hashes of loaded rule files or matching identifiers before and after compaction, run first; model-based semantic checks are kept separate. It maps seven security controls to those records: least privilege, durable authorization, information flow control, monitors and guardrails, accountability, failure attribution and recovery.
The evaluation is the strongest evidence for the thesis. Across 240 multi-step trajectories (120 benign, 120 fault-injected), a log of tool calls alone caught 16.7% of the injected multi-transfer violations, while the full transfer-audit schema caught 98.3%, both with a 0% false-positive rate in that test (arXiv:2611.23346).
Approvals That Do Not Survive the Harness
A separate preprint, submitted on November 14, asks whether human approval actually controls what happens: rejected work must stop and approved work must complete (Deng et al., arXiv:2611.19237). It calls these properties refusal safety and bounded authorization completeness, and audits three boundaries: the model selecting a call, the harness scheduling and resuming it, and the workflow routing the decision to a protected effect. The gaps it reports sit at different layers:
- Model layer: the paired model panel left 7 of 300 approvals unfinished, with zero denied effects.
- Harness layer: Google ADK's parallel confirmation lost an approved branch in three tasks per model.
- Harness layer: on restoration, CrewAI 1.15.37 omitted 336 of 672 approved effects, even though the decisions and business state were preserved.
- Workflow layer: 4 of 59 policy-comparable public n8n workflows executed the protected effect after Reject.
Only the restoration and resume failures are state-transfer failures in the strict sense, and they are the ones a tool-call log cannot explain: the approval was recorded, but the work it released never ran. The paper's answer is an audit contract that checks intent, generated calls, dispatch and native effects against each other.
Measuring What Controls Actually See
guardscope, submitted on November 11, measures which agent actions a runtime control can check at all (Toyoda et al., arXiv:2611.15223). It predicts from source code which of an agent's 15 action paths a control covers, then runs a workload on one of eight topology designs and classifies every executed action as checked, bypassed or prevented against kernel-level ground truth. The authors ran it with two agent harnesses and real controls, including Microsoft's agent governance toolkit and agentgateway. The paper is about action paths, not state transfers, but it makes the same point from another side: having a gateway is not evidence that the gateway sees every action.
Policy Enforcement Without an LLM Judge
Tether, submitted on November 19, runs each proposed agent action in an ephemeral sandbox, infers what the action did from observed system calls, and enforces stateful policies written in a Lean 4 DSL (Sistla et al., arXiv:2611.25599). Writing policies in Lean lets the authors formally verify that they satisfy global invariants, such as no network access after reading a sensitive file. On coding-agent benchmarks, a single shared policy blocked 100% of in-scope attacks with a benign false-positive rate under 5%. Tether avoids relying on an LLM monitor, which the authors note lacks provable safety guarantees.
What a Production Audit Trail Has to Hold
A September 1 guide from Cogniver, a vendor of AI approval workflows, draws on guidance from MightyBot, ARMO, ibl.ai, Nylas CLI and Collibra and lists ten requirements to "treat as a production control specification" (Cogniver): stable identifiers, every relevant actor, pinned versions, the preserved trigger, data lineage, every condition evaluation, every invocation, external effects, ordered UTC timestamps, and records of human control. Separately, it recommends tamper evidence and alerts when expected events disappear or logging is disabled. Read against the state-transfer research, the hard part is keeping those identifiers, versions and approval records intact when context is compacted or a run is restored.
Where Vendors Are Building
Some shipping systems already treat paused-and-resumed state as a first-class object. HubSpot says the transcript in its Aviator harness "is also the complete state of a conversation": it serializes to JSON, can be checkpointed and resumed on a different machine, and lets an agent pause while it waits for a human to approve a tool call (HubSpot Engineering). Microsoft open-sourced Agent Hooks and the Agent Control Specification (ACS), an open request-and-decision contract that gathers evidence through annotators (safety classifiers, LLM judges, deterministic rules or any external API) and passes it to a policy engine such as Rego; they underpin its managed Azure service, agent controls (Microsoft). Docker AI Governance adds network and filesystem policies around microVM sandboxes, Cedar policies for MCP activity through Docker's MCP Gateway, and structured audit records that can be forwarded to SIEM workflows (Docker). Sysdig says its AI Defense pairs a kernel-level record with harness instrumentation, and that for every action an agent writes in its log, the kernel sees about seven processes doing the work (Sysdig).
The frameworks move fast too. Between November 13 and November 20, CrewAI (1.15.44), LangGraph (1.2.24), OpenAI Agents Python (v0.25.0), Pydantic AI (v2.71.0), Google ADK Python (v2.15.0) and LlamaIndex (v0.14.26) all shipped releases. We have not checked whether any of them changes the restoration or parallel-confirmation behaviour measured above; the CrewAI result was on 1.15.37.
Hlinor's View
Hlinor's view: Agents are long-running stateful processes, and governance that only watches tool calls is like auditing a database by reading the query parser while ignoring the transaction log, the checkpoint and the rollback. The evidence so far is a handful of preprints and vendor posts, not a settled field, but it points one way. Teams should test whether an approval survives a pause, a compaction and a restore, and they should ask harness vendors for hash-chained records at those boundaries. We expect buyers to start asking for exactly that in governance reviews. After every framework upgrade, re-run the approve-then-restore test.
Sources
- Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness (Deep Shah, Nehal Kathrotia), arXiv, 2026-11-17, https://arxiv.org/abs/2611.23346
- Auditable Authorization for LLM Agents: Empirical Evidence Across Model, Harness, and Workflow Layers (Bangjie Deng et al.), arXiv, 2026-11-14, https://arxiv.org/abs/2611.19237
- guardscope: Measuring Which Actions Runtime Controls Check in LLM Agent Systems (Kentaroh Toyoda, Alex Leung, Bhuvaneswari U), arXiv, 2026-11-11, https://arxiv.org/abs/2611.15223
- Tether: Verified Stateful Policy Engine for System-Level Monitoring of Agents (Meghana Aparna Sistla et al.), arXiv, 2026-11-19, https://arxiv.org/abs/2611.25599
- AI Agent Audit Trail Requirements for Controlled Operations, Cogniver, 2026-09-01, https://cogniver.com/insights/ai-agent-audit-trail-requirements
- Aviator: HubSpot's Agent Harness, HubSpot Engineering, 2026-11-19, https://hubspot.engineering/blog/aviator-hubspots-agent-harness/
- Agent controls: An enterprise control plane for agent behavior, Microsoft, 2026-11-18, https://commandline.microsoft.com/introducing-agent-controls-an-enterprise-control-plane-for-agent-behavior/
- Securing Agentic AI with Docker AI Governance, Docker, 2026-11-18, https://www.docker.com/blog/agentic-ai-security-docker-governance/
- Sysdig Launches Sysdig AI Defense, Bringing Runtime Security to Every Action AI Agents Take, Sysdig, 2026-11-04, https://www.sysdig.com/press-releases/sysdig-launches-sysdig-ai-defense-bringing-runtime-security-to-every-action-ai-agents-take
- CrewAI 1.15.44, GitHub, 2026-11-19, https://github.com/crewAIInc/crewAI/releases/tag/1.15.44
- LangGraph 1.2.24, GitHub, 2026-11-20, https://github.com/langchain-ai/langgraph/releases/tag/1.2.24
- OpenAI Agents Python v0.25.0, GitHub, 2026-11-18, https://github.com/openai/openai-agents-python/releases/tag/v0.25.0
- Pydantic AI v2.71.0, GitHub, 2026-11-19, https://github.com/pydantic/pydantic-ai/releases/tag/v2.71.0
- Google ADK Python v2.15.0, GitHub, 2026-11-19, https://github.com/google/adk-python/releases/tag/v2.15.0
- LlamaIndex v0.14.26, GitHub, 2026-11-13, https://github.com/run-llama/llama_index/releases/tag/v0.14.26