Read the fallback config as a latency budget. The primary `customer_api.search` gets a 30-second timeout and two retries, `cache.lookup` gets five seconds and none, and the last hop is a template message [10]. If `retries: 2` means two attempts after the first, the chain can burn 30 seconds three times plus five before the user sees the template, which is 95 seconds [1]. A 95-second wait to be told the system cannot help is its own incident class. The shape of that config is right. The numbers in it are a workload assumption, and you have to re-derive them against your own dependency latencies.
The gate is two vetoes on one object: `should_respond` returns False when `response.confidence` is under the threshold, and again when `response.missing_context` is set [5]. Look at what neither condition inspects. One of the failures in the author's log is an agent that produced correct output and attributed it to the wrong source document [9]. That response carries high confidence and no missing context, so it passes both checks and ships. An uncertainty channel catches absence of knowledge. Misplaced provenance needs a separate check against the retrieved span.
The 60 percent [8] is a claim about one incident population, and the author reports it from agents doing customer support, data extraction and workflow automation for enterprises [19]. For the number to move, two things have to hold. Your incidents have to be dominated by confident-wrong answers rather than tool failures or stale retrieval, and you need the held-out, human-labelled set the threshold was calibrated against [7]. The signal has a bill attached too. Token-level entropy assumes your endpoint hands back per-token probabilities, and self-consistency means re-running the same query several times and measuring variance, which multiplies inference cost on every request including the ones that end in a refusal [6].
There is a seam between the two snippets worth noticing. `AgentTrace` stores `confidenceScores` as a keyed map alongside a list of uncertainty snapshots [17], while the gate reads a single scalar [5]. Somewhere between those two representations is an aggregation rule, and that rule decides whether the threshold means anything.
The strongest part of the post is not the gate. It is the admission that isolated evals looked perfect and production did not [13], and that the difference only showed up once the harness modelled a cold cache, 0.3 latency jitter, 50 concurrent users and injected dependency failures [14]. That harness is what reclassified 30 percent of the errors as timeout cascades, where a call timed out, retried, and the retry added to the load that caused the timeout [16]. No prompt reaches that. Circuit breakers do.
