Silent Tool Failure: The AI Failure Mode the Literature Missed

An AI agent runs a command. The command errors. The agent keeps going, the work comes back looking finished, and the human accepts it. No alarm fires, because nothing in the interface treats this as a failure.

That pattern kept showing up in our data, and the research literature had no name for it. This is how invisible AI failures work: the dangerous ones do not look like failures at all. We named the pattern, measured it, and it is the first finding to clear our publication gate.

The short version: Research on 100,000 real chatbot conversations (Potts and Sudhof, Stanford/Bigspin) found that 79% of AI failures leave no visible trace, and ordinary monitoring catches about 12% of them. That research mapped eight failure archetypes. When we instrumented real agentic work rather than chatbot conversations, the eight were not enough: a ninth kept appearing. A tool call errors, the agent proceeds, the human never addresses it. We published it as silent_tool_failure, gated at inter-annotator reliability of Cohen's κ 0.774 against our 0.70 publication bar.

The problem: failure that arrives looking finished

Three findings define why this research exists. First, most AI failure is invisible: no error message, no complaint, no correction, in 79% of cases. Second, humans adopt AI output without engaging their own reasoning, and their confidence rises even when the output is wrong (Shaw and Nave at Wharton, across 1,372 participants). Third, the failure corrupts its own alarm: repeated AI use recalibrates the sense that something is off to match the machine's confidence.

That is a profile no one can self-diagnose, because the instrument you would use to catch it is the thing being rewired. It is why the interaction has to be instrumented from outside, which is what our cognitive surrender research program does.

What we did

We instrumented our own work first. Subject-Zero is Ivanooo's founder, and the corpus is his real agentic sessions: full transcripts, not surveys, not lab tasks.

The scale surprised us before the finding did. Of 22,000 turns in the transcripts, only 537 were genuinely human. An assistant-to-human ratio near 24 to 1 is the shape of agentic work itself: the machine acts hundreds of times per human glance.

The method rests on one move. You cannot plant errors in someone's real work, so real work looks unmeasurable. But the known flaw catalogue is the ground truth: if you know what the characteristic failures look like, you know what failing to catch one looks like, without planting anything. Episodes get labeled signals-first, and nothing publishes below Cohen's κ 0.70. Our first labeling attempt failed that bar at 0.608, and repairing the disagreement is what exposed the finding.

The finding

The chatbot archetypes describe conversation failures: the confidence trap, the silent mismatch, drift, the death spiral. Agentic work adds something conversations do not have: tool calls, where the machine acts on the world. That is where the ninth archetype lives.

silent_tool_failure: a tool call errors, the agent proceeds without resolving it, and the operator moves on without addressing it.

Where we looked (Subject-Zero corpus)	Share of failures that were invisible
Dialogue: the human-AI conversation	12%
The agent's tool loop	Near 100%

The asymmetry is the point. In dialogue, Subject-Zero caught most of what went wrong, because dialogue is where a human's attention lives. Inside the tool loop, failures passed unaddressed almost every time. He was watching the conversation, not the agent.

Why it matters

Agentic AI multiplies tool calls per human decision, so it multiplies exactly the surface where failure is least visible. Each silently accepted error feeds the chain our research maps: acceptance becomes surrender, surrender becomes capability decay, and the work converges toward whatever the machine produced. The Darkness Library catalogues the flaws; this finding shows where they pass uncaught, and the decay only shows across the body of work, never in one session.

Two honesty rails, stated plainly. This is a one-subject study: it establishes that the archetype exists and can be measured reliably, not how prevalent it is across operators. And we score observed signals, never capability: the data says what happened in the work, not what anyone is capable of.

See your own

The instrument that produced this finding runs on a single conversation. Paste one AI conversation and get a reading: are you directing the machine, or has it started directing you? For teams running agentic production, the same instrumentation runs continuously as an observability layer.

FAQ

What is silent tool failure? A failure archetype in agentic AI work: a tool call errors, the agent proceeds without resolving it, and the human operator moves on without addressing it. No alarm fires because the work still arrives looking finished.

What does it mean that the finding is "gated"? Two independent annotators label the same episodes, and we compute agreement beyond chance (Cohen's κ). Nothing publishes below κ 0.70. This finding cleared at 0.774.

Is the 79% invisible-failure figure yours? No. It comes from Potts and Sudhof's analysis of 100,000 real chatbot conversations. Our contribution is the ninth archetype and the dialogue-versus-tool-loop asymmetry, measured on real agentic work.

Can a one-subject study prove anything? It proves existence, not prevalence: the archetype is real, recurrent, and measurable at research-grade reliability. Prevalence across operators is the next study, and the corpus is growing.

How do I know if this is happening in my work? You will not feel it, because the failure is invisible from the inside by definition. Score a real conversation, or instrument the work. Introspection is the one tool this failure disables.