Silent Semantic Drift -- Inter-Agent Series # 3

Ming's Substack

SubscribeSign in

Silent Semantic Drift -- Inter-Agent Series # 3<br>The Most Dangerous Agent Failure Looks Exactly Like Success.

Ming Chiu<br>May 20, 2026

Share

The most expensive agent failure of the next several years won’t look like a failure at all. It’ll look like success — same green dashboard, same 200 OK, same tidy log entry on both sides of the wire.<br>Except the two sides have just agreed to two different things, and nobody will know for three weeks.<br>That failure has no name in your incident vocabulary. It’s not in your CISO’s playbook. It doesn’t pattern-match to anything your QA team has tested for. When it surfaces — at delivery, at reconciliation, at the lawyer’s office — the audit trail will swear nothing went wrong.<br>I’ve been calling this silent semantic drift. I think you may need the term soon.

What it is

Silent semantic drift is the gradual, undetected divergence between what two or more agents believe they have agreed upon over the course of a multi-turn exchange.<br>It requires no bad actor. It produces no error message. It is the natural consequence of two or more systems exchanging natural language while each side resolves that language privately, against contexts the other side cannot see.<br>The failure isn’t happening inside either agent — it’s happening between them, and that’s the thing existing tools were not built to see.<br>It also lives in one specific place: the interaction phase of the full-cycle agentic experience, the long, ambiguous middle that sits between encounter and settlement — the back-and-forth where a quote becomes an order and terms get pinned down (or don’t). The phase your current stack is mostly silent about.<br>Why silent

Two weeks ago I argued that a 200 OK is a false positive for alignment. Drift is what fills that gap. Each individual exchange parses as fine — headers well-formed, schemas valid, confirmations coming back in the affirmative. Watch it on a console and every turn looks healthy.<br>The conventional debugging instinct — something must be erroring; let me find the error — fails here, because nothing is erroring. The transaction completes, settlement fires, the dashboard stays green. Most teams I talk to don’t have it on their list yet.<br>Why semantic

The drift isn’t in the bytes. It’s in the meaning.<br>Consider a request like “send the standard concentration.” Both agents parse the sentence. The word standard, however, gets resolved on each side against a different reference: one agent’s product catalog says one thing; the other’s says another. Each side is internally consistent. Neither is hallucinating. The disagreement is entirely about what real-world referent the words are pointing at — and there is no shared shelf the two agents can point at to disambiguate.<br>This is what psycholinguists since Herbert Clark have called the common ground problem: two parties communicating without a continuously verified shared frame of reference will diverge in their interpretations. The active process by which humans maintain that shared frame has its own name — grounding — and it runs in the background of every successful conversation, mostly through small confirming behaviors: eye contact, nods, repetition, “so what you’re saying is…” Almost none of that crosses the wire between agents, by design.<br>JSON, as I put it last time, is a courier, not a referee. Silent semantic drift is what happens when nobody refs.

Why drift

A single semantic mismatch on a single turn isn’t, by itself, catastrophic. The danger is what happens across turns. Each subsequent message is interpreted through the already-divergent frame, which means clarifications intended to resolve the divergence often deepen it. The buyer’s agent asks about delivery timeline; the seller’s agent quotes a timeline for the wrong SKU; the buyer’s agent confirms; the divergence is now load-bearing for the rest of the transaction. By turn six, the agents aren’t negotiating the same deal. By turn ten, the gap is wide enough that you’d see it in seconds — if anyone were looking.<br>Researchers studying multi-agent LLM systems have started cataloguing this empirically. A 2025 taxonomy effort annotated more than 1,600 failure traces across seven multi-agent frameworks and identified inter-agent misalignment as one of three dominant failure categories — and concluded these were fundamental design flaws, not artifacts of specific systems. Separately, benchmarking work has shown that LLM agents negotiating across model families produce measurably worse deal outcomes than agents within the same family.<br>The classical analogy is the bullwhip effect in supply chains, where small upstream misreadings amplify into large downstream distortions. The shape is similar; the substance isn’t. Bullwhip distortions are about quantities — how much, how many, how soon. Drift is about qualities — what standard refers to, what the order covers, what delivered counts as. Same...