The agent that fires too often becomes the agent you ignore
You have a smoke detector in your kitchen. It goes off every time you make toast. The first time it fires, you jump up and check for fire. The second time, you wave a towel at it. By the fifth time, you pull the battery out. Three months later, an actual grease fire starts on the stove. The detector is sitting in a drawer, battery beside it.
This is not a story about negligence. It is a story about a system that trained you — through repeated false activations — to treat real danger as noise. Your cognitive agents carry the same risk. Every time an agent fires when it should not, it does not just waste your attention in that moment. It degrades the trust relationship between you and the system, making it incrementally more likely that you will ignore the agent when it fires correctly.
The false positive rate measures exactly this: how often does an agent activate when the trigger condition is not actually present? It is the ratio of false activations to total activations, and it is one of the most consequential metrics in your entire monitoring infrastructure.
Signal detection theory: separating sensitivity from bias
The formal framework for understanding false positives comes from signal detection theory (SDT), developed in the 1950s for radar operators who needed to distinguish enemy aircraft from flocks of birds on noisy screens. The core insight is that every detection system — radar, medical test, cognitive agent — faces the same four possible outcomes:
| | Signal present | Signal absent | | ---------------- | --------------------- | --------------------------------- | | Agent fires | Hit (true positive) | False alarm (false positive) | | Agent silent | Miss (false negative) | Correct rejection (true negative) |
Two independent factors determine where your agent lands in this matrix. Sensitivity (measured as d-prime in SDT) captures how well the agent can discriminate signal from noise — how different the trigger condition looks from background activity. Criterion (measured as beta or c) captures the threshold at which the agent decides to fire — how much evidence it requires before activating.
Here is the critical insight: you can have a highly sensitive agent that still produces an overwhelming number of false positives, simply because its criterion is set too low. The agent can tell signal from noise reasonably well, but it fires at the slightest hint of signal, catching everything real along with everything that merely resembles the real thing.
A liberal criterion — one that fires easily — minimizes false negatives at the cost of maximizing false positives. A conservative criterion does the reverse. You cannot minimize both simultaneously. This is the fundamental tradeoff, and understanding it prevents you from thinking a noisy agent is simply broken. It may be well-calibrated for sensitivity but poorly calibrated for threshold.
What happens when the rate climbs: the cry wolf effect
Shlomo Breznitz, in his 1984 book Cry Wolf: The Psychology of False Alarms, conducted experiments where participants were repeatedly warned that a painful shock was imminent — and then the shock was canceled. Each false alarm reduced participants' physiological stress response to the next warning. The body learned, through conditioning, that the warning did not predict the outcome. By the fifth or sixth false alarm, participants showed minimal autonomic arousal — their heart rate, skin conductance, and cortisol barely moved. The warning system had trained them into non-response.
This is not a metaphor for what happens to your cognitive agents. It is the mechanism. When your "review your priorities" agent fires every morning but the trigger condition (a significant change in project status) only applies twice a month, your brain habituates. The notification becomes wallpaper. Breznitz's research demonstrated that this habituation is not a conscious choice — it is a neurological adaptation. You do not decide to stop caring about the agent. Your nervous system decides for you, because it has learned that the signal is not predictive.
The Aesop's fable version of this research — the boy who cried wolf — captures the social dimension. Maureen O'Sullivan's research on the "boy-who-cried-wolf effect" showed that repeated false claims erode not just trust in specific statements but trust in the source itself. Applied to your cognitive agents: an agent that consistently false-fires does not just lose credibility for that particular trigger. It contaminates your trust in the entire monitoring system. You start treating all agent notifications as probably-noise, including the ones from agents that are well-calibrated.
Alert fatigue: the empirical evidence from medicine and DevOps
Healthcare provides the starkest evidence of what unchecked false positive rates produce.
Between 72 and 99 percent of clinical alarms in hospitals are false. Patients in intensive care units are exposed to as many as 700 physiologic monitor alarms per patient per day. The Joint Commission's sentinel event database recorded 98 alarm-related events between 2009 and 2012, of which 80 resulted in patient death. From 2005 to 2010, 216 U.S. hospital patients died in incidents related to management of monitor alarms. The ECRI Institute listed alarm fatigue as the number one health technology hazard for four consecutive years starting in 2012.
The mechanism is identical to Breznitz's laboratory findings, scaled to life-and-death stakes. Clinicians who experience hundreds of non-actionable alarms per shift develop desensitization. They silence alarms at the central station without checking the patient. In some documented cases, they permanently disable alarm systems. The false positives did not just waste attention — they removed the infrastructure that was supposed to catch true positives.
In DevOps, the pattern repeats. Teams receive over 2,000 alerts weekly, with only about 3 percent needing immediate action. In cybersecurity, one survey found that 52 percent of alerts were false and 64 percent were redundant. PagerDuty's 2021 survey found that most incident responders received over 10 alerts per shift, most of which could not be acted on. A 2025 Catchpoint report found nearly 70 percent of site reliability engineers said on-call stress — driven largely by alert noise — directly contributed to burnout and attrition.
These are not engineering problems. They are false positive rate problems. The monitoring infrastructure exists. The sensors work. The threshold is wrong.
Precision: the metric that measures false positive damage
In machine learning classification, the metric that directly captures false positive damage is precision:
Precision = True Positives / (True Positives + False Positives)
An agent with 90% precision fires 10 times and is correct 9 of those times. An agent with 30% precision fires 10 times and is correct only 3 — meaning 7 out of 10 activations are false. You can see immediately why precision below 50% is dangerous: the agent is wrong more often than it is right, and your lived experience of the agent becomes one of mostly-incorrect activations.
Precision is the complement of false discovery rate. If your precision is 40%, your false discovery rate is 60% — meaning 60% of the time the agent fires, it is discovering nothing real.
This is distinct from false positive rate in the technical sense (false positives divided by total negatives), but for cognitive agent monitoring, precision is the more actionable metric because it answers the question you actually care about: when this agent fires, how likely is it to be right?
Track precision, not just activation count. An agent that fires 50 times a week with 90% precision (45 true positives, 5 false) is dramatically more useful than an agent that fires 50 times a week with 20% precision (10 true positives, 40 false). The second agent is not twice as noisy — it is an active threat to your attention infrastructure, because it is training you to stop listening.
Why false positives are costlier than they appear
The direct cost of a false positive is the attention spent evaluating and dismissing it — typically seconds. This is why people underestimate the damage. Seconds seem cheap.
But the indirect costs compound:
Trust erosion is nonlinear. The first false positive barely registers. The tenth creates mild annoyance. The fiftieth triggers the response "I should probably turn this off." Trust does not degrade linearly with false positive count — it follows a threshold function where a critical mass of false activations causes a sudden, discontinuous drop in compliance. Breznitz's experiments showed this pattern clearly: participants' stress responses did not decline smoothly — they dropped sharply after a small number of false alarms.
Attention is a shared resource. Every false positive that interrupts your focus state does not just cost the seconds of the interruption. Research on interruption recovery consistently shows that returning to a deep-focus task after an interruption takes 10 to 23 minutes, not the 5 seconds the notification itself consumed. A false positive during a flow state costs vastly more than a false positive during idle time.
Cross-agent contamination. As O'Sullivan's research demonstrated, repeated false alarms erode trust in the source, not just the specific claim. If your "check your energy level" agent false-fires constantly, you begin discounting your "check your calendar" agent too — even if the second agent has perfect precision — because the notification channel itself has become associated with noise.
Learned helplessness. At extreme false positive rates, you do not just ignore the agent — you stop believing that your monitoring system can work at all. The failure of one poorly calibrated agent becomes evidence that cognitive agents in general are not useful, the same way that a colleague who is consistently unreliable makes you skeptical about delegation as a strategy, not just delegation to that specific person.
Tuning the threshold: how to reduce false positives without losing true positives
The signal detection theory tradeoff means you cannot simply "remove false positives." Tightening the trigger condition to eliminate false positives will simultaneously increase false negatives — the agent will miss real instances of the condition it is supposed to detect. The goal is not zero false positives. The goal is a precision level that sustains trust.
Three approaches:
Narrow the trigger condition. If your "take a break" agent fires whenever your heart rate exceeds 90 bpm, you are using a trigger that conflates exercise, laughter, excitement, and actual stress. Add context: heart rate above 90 bpm AND you have been seated for more than 20 minutes AND it is during working hours. Each additional condition acts as a filter that removes an entire category of false positives.
Add a confirmation step. Instead of firing on a single signal, require the trigger to persist for a defined period. An elevated heart rate for 3 seconds is noise. An elevated heart rate for 10 minutes while seated is a meaningful signal. Debouncing — requiring the condition to hold for a duration before the agent activates — is the simplest and most effective false positive reduction technique.
Raise the threshold, then monitor false negatives. Move the criterion more conservative and watch whether the agent starts missing real events. If you raise the heart rate threshold from 90 to 110 bpm and the agent still catches every genuine stress episode, you have found free precision — false positives you were paying for that had no corresponding benefit. This is why L-0549 (false negative rate) is the necessary complement to this lesson: you cannot tune one without monitoring the other.
The bridge to false negatives
Every adjustment you make to reduce false positives shifts the balance toward potential false negatives. A more conservative trigger misses less obvious instances of the real condition. A narrower trigger definition excludes edge cases that may be genuine. A debounce window means the agent cannot catch rapid-onset events.
This is not a reason to avoid tuning. It is a reason to monitor both metrics simultaneously. The false positive rate tells you how much trust the agent is burning. The false negative rate — which is the subject of the next lesson — tells you how much signal the agent is missing. Together, they define the operating characteristic of your agent: the curve that maps every possible threshold setting to its corresponding pair of error rates.
An agent you can trust is not an agent with zero errors. It is an agent whose errors you understand, whose error rates you measure, and whose threshold you have deliberately set based on the relative cost of each error type in your specific context.
The question is not "does this agent fire correctly?" The question is "when this agent fires, do I still believe it?"