The procurement question your AI governance vendor cannot answer
Ask your AI governance vendor what they call the failure mode where agents start drifting off-topic over the course of a multi-agent run. If they cannot name it, they cannot detect it.
That is the question regulated buyers should be asking in 2026, and almost no vendor has a clean answer. Most reach for "hallucination", the wrong category. Others say "guardrail violation", which tells you nothing about how the intent of the system moved over time.
The failure mode has a name. It is topic adherence drift, the defining production risk of agentic AI.
What topic adherence drift actually is
In a single-agent LLM system, the input maps to the output in one call. Most governance tooling was built around that architecture.
Multi-agent systems do not behave that way. When agents call each other, the working intent moves with every hand-off. The agent that started routing a customer query ends up giving financial advice the bank never authorised it to give.
This is topic adherence drift. It is a property of agentic architecture, not a bug in any individual model. Each agent reinterprets the context it inherits, and small reinterpretations compound. By the fifth hand-off, the system is operating on a working intent the original prompt did not contain.
The behaviour is statistical. It can be measured and evidenced under audit, but only by a platform with a name for it.
Why no category means no question, and no question means no test
Regulation moves through language. The EU AI Act under Article 9 requires continuous risk management across the lifecycle. Article 72 requires post-market monitoring. Neither names topic adherence drift, because the category did not exist when the text was written.
Practitioners can describe what goes wrong. They cannot name it. The regulator conversation defaults back to hallucination and prompt injection, and the real failure mode runs uncatalogued. This is Agentic Theatre: the architecture diagram says "multi-agent governance", the runtime evidence is single-call hallucination scoring, and the gap only shows up when a regulator asks how a decision propagated across the chain.
The buyer who can name the failure mode in their RFP forces vendors to demonstrate detection, not promise it.
How Disseqt monitors topic adherence drift
Disseqt runs 65 input validators across four families: base, RAG, agentic, and MCP. The agentic family is purpose-built for failure modes that only appear in multi-agent systems, and topic adherence drift sits at the centre of it.
The agentic validators monitor the intent vector across hand-offs. Each agent-to-agent call produces a measurable signal: how far the working intent has moved from the parent call, and from the original system prompt. The validator scores that movement against thresholds the enterprise sets per workflow. When drift crosses threshold, the platform escalates, routes for human review, or blocks the next hand-off.
This is runtime monitoring, not pre-deployment testing. It lives in Protect & Enforce, the pillar of the AI Assurance Lifecycle between the agent and production. Test & Detect catches what you can simulate before go-live. Protect & Enforce catches what only appears in the field.
The validators run inline at sub-50ms, model-agnostic. Every hand-off is logged with a timestamped intent vector score, the threshold, the policy, and the action taken, reproducible under audit.
That is what Article 9 requires when the architecture is agentic, and what Article 72 post-market monitoring requires when the system is multi-agent. The FCA, the SEC, and ISO/IEC 42001 will read deployments against the same standard.
Bottom Line
Topic adherence drift is the failure mode the market does not yet have language for. Single-call testing does not catch it. PowerPoint Governance cannot evidence it. The vendor that cannot name it has no validator family built to detect it.
Disseqt names the failure mode, monitors it inline with the agentic validator family, and surfaces it in Protect & Enforce. One platform, three pillars of the AI Assurance Lifecycle, sold as the Assurance Layer for Enterprise AI. The architecture that answers the regulator's question is the one that could name it first.
FAQs
What is topic adherence drift in agentic AI?
Topic adherence drift is a measurable shift in the working intent of a multi-agent system across agent-to-agent hand-offs. Each agent reinterprets the context it inherits, and small reinterpretations compound. It is a property of agentic architecture, not a model bug.
How is topic adherence drift different from hallucination?
Hallucination is a single-call failure: output not grounded in input or training data. Topic adherence drift is a multi-call failure: the working intent of the system moves across hand-offs, even when no individual call is hallucinating. Hallucination scoring will not catch drift.
Why does topic adherence drift matter for EU AI Act compliance?
Article 9 requires continuous risk management across the lifecycle, and Article 72 requires post-market monitoring. In agentic systems, that means monitoring how the intent of the system moves between calls, not just whether individual calls breached a guardrail. A stack that cannot evidence drift detection cannot evidence the continuous obligation.
Where does drift monitoring sit in the Disseqt platform?
Drift monitoring lives in Protect & Enforce, the runtime pillar of the AI Assurance Lifecycle. The agentic validator family scores intent movement inline at sub-50ms, escalates or blocks on threshold breach, and surfaces drift in the agentic observability layer. The evidence record feeds Prove & Comply for audit.
AUTHOR
Apoorva Kumar
CEO and Co-Founder
Apoorva Kumar is Founder and CEO at Disseqt, where he's building the assurance layer for enterprise agentic AI. Previously Senior Manager of Product Management at Microsoft — leading Teams and SharePoint Premium and at AWS, where he built and shipped severless compute for high-performance workloads
