--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/42/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/42/foundations" issue: 42 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Sevda Polat","Sable Whitford"] published: 2026-08-17 updated: 2026-08-13 articles: 24 --- # TinyFish > A journal for living in the agentic age [Read on the website]() ## Issue 42 — Foundations Conceptual clarity earned from building at scale ### Review as Control - [The Review That Isn't There]() — By [Sevda Polat]() · Published 2026-08-12 · Format: Review as Control A survey of 86 deployed agent systems found that 74% use human verification to ensure quality. None of them compared outcomes with the review step against outcomes without it. Separately, a meta-analysis of 106 human-AI experiments found that adding a human reviewer to a stronger AI performer tends to make the result worse. The step is staffed, budgeted, and cited whenever regulators ask how quality is maintained. Whether it improves anything remains, in most organizations, untested. ### Designing Effective Review - [Designing Review That Resists Its Own Collapse]() — By [Sevda Polat]() · Published 2026-08-12 · Format: Designing Effective Review An FDA-authorized pathology tool requires the doctor to record a diagnosis before the AI reveals its own. Commit first, then compare: that ordering contains most of what you need to know about designing review that resists rubber-stamping. Three areas of design detail decide whether a checkpoint holds: what evidence the reviewer sees, when review gets triggered, and how you find out whether any of it is still working. ### Adjacent Thinking - **Human-AI Teams Usually Underperform the Better Solo Player** — [Source]() — Published 2026-08-12 A meta-analysis of 106 studies finds most human-AI pairings lose to whichever performer was already stronger alone. - **When the Model Grades Its Own Homework** — [Source]() — Published 2026-08-12 LLM suggestions shift annotator labels toward the model's own outputs, inflating the evaluated model's reported performance. - **A Safety Engineering Vocabulary for Agent Oversight Timing** — [Source]() — Published 2026-08-12 STPA's four unsafe control action types offer formal language for why correctly specified checkpoints still fail. - **Alert Fatigue Degrades Review Quality Across the Board** — [Source]() — Published 2026-08-12 Over half of SOC analysts report overwhelm, and teams resolve fewer than half their alerts per shift. - **Longer AI Explanations Raise Confidence Without Improving Accuracy** — [Source]() — Published 2026-08-12 People evaluating LLM explanations performed near chance at telling correct answers from incorrect ones. - **The Out-of-the-Loop Problem Hasn't Gone Away** — [Source]() — Published 2026-08-12 Passive monitoring erodes the situation awareness, skill retention, and vigilance that meaningful oversight depends on. ### Review Failure Modes Human review is the go-to safeguard when agents handle consequential work. But the review step has its own failure modes, well-documented in fields like clinical alerting and code review, and increasingly visible in agent workflows. Research points to at least six. The reviewer might arrive too late, see only what the agent chose to surface, or lack real authority to intervene. Cognitive anchoring and sheer approval volume degrade attention from the inside. And without outcome feedback, reviewers never learn to calibrate. These failures interact, and they tend to accumulate without drawing attention to themselves. - **The Reviewer Shows Up After the Window Closes** — [Source]() — Published 2026-08-12 Tag: Late Arrival A Taobao field study found that when human agents took over after customer frustration had already built, chat duration rose 40.8% and satisfaction dropped nearly a full point compared to matched human-handled cases. By that point, the viable options had already narrowed. - **Reviewers See the Agent's Summary, Not the Situation** — [Source]() — Published 2026-08-12 Tag: Framing Bias A Nature Machine Intelligence study found people evaluating LLM explanations performed barely above chance at distinguishing correct from incorrect answers. Longer explanations raised confidence without improving discrimination. The explanation shaped the reviewer's judgment rather than informing it. - **Review Without Authority to Act Is Just Observation** — [Source]() — Published 2026-08-12 Tag: Authority Gap When a reviewer can see a decision but cannot reject, pause, or reroute it, the oversight is ceremonial. A common pattern: approval requests that time out into automatic execution. A position paper warns this generates false legitimacy without actual control. - **The Agent's Confidence Anchors the Reviewer's Own Judgment** — [Source]() — Published 2026-08-12 Tag: Confidence Anchoring An AAAI experiment found participants accepted high-confidence AI advice more often, even when it was wrong. Well-calibrated confidence improved accuracy by 20%, but miscalibrated confidence produced gains indistinguishable from zero. Anchoring sets in before the reviewer forms an independent view. - **Too Many Items and the Reviewer Stops Noticing** — [Source]() — Published 2026-08-12 Tag: Approval Fatigue Computerized prescribing alert systems show override rates from 46% to 96%. Code review experiments find defect detection dropping from 59% to 35% as patch complexity grows. At high volume, review becomes pattern-matching against a default of "approve." - **Reviewers Never Learn Whether Their Approvals Were Right** — [Source]() — Published 2026-08-12 Tag: Missing Feedback An ACL 2025 study found model suggestions shifted annotator labels toward the AI's output while raising self-reported confidence. When those labels became evaluation data, reported model performance inflated. Without outcome feedback, "human-approved" describes a process step, not a quality signal. ### Review Design Patterns Human review is the default answer to "how do we keep agents from doing something wrong?" But review quality depends heavily on how it's structured. A reviewer who sees the agent's recommendation before examining the evidence is already anchored. One who shares the agent's full context may miss systematic drift. One who arrives after the meaningful choices have closed adds a step without adding a check. Six design patterns target specific failure modes in review workflows. Some protect reviewer independence, others restructure what the reviewer sees and when, and a few govern how long an agent's authority persists. Each costs something, and each degrades differently under volume. - **Blind Review Catches Drift First Reviewers Will Miss** — [Source]() — Published 2026-08-12 Tag: Blind Review A second reviewer evaluates the work without seeing the agent's output or the first reviewer's assessment. The goal is detecting systematic drift, not duplicating effort. Production workflows typically sample 10-15% of items, selecting cases likely to expose correlated errors rather than random variation. - **Seeded Disagreements Show Whether Reviewers Catch Real Errors** — [Source]() — Published 2026-08-12 Tag: Seeded Disagreements Known-answer cases get quietly inserted into a reviewer's queue. Performance on these planted items reveals whether the reviewer is catching errors or waving things through. The ongoing cost is maintaining a gold standard that tracks the agent's current failure modes, not last quarter's. - **Showing Evidence Before Recommendations Weakens the Anchoring Effect** — [Source]() — Published 2026-08-12 Tag: Evidence First When reviewers see the AI recommendation first, it shapes how they read the supporting evidence. Withholding the recommendation until the reviewer forms an initial judgment reduces this anchoring. Research found reviewers were 65% more likely to revise answers in a human-first workflow. - **Expiring Permissions Force Agents to Re-Earn Their Authority** — [Source]() — Published 2026-08-12 Tag: Bounded Authority Agent permissions should expire and require renewal rather than persist by default. Sensitive operations may warrant token lifetimes of 60-300 seconds. The harder design question: defining validity conditions beyond clock time, like price changes or shipment status, and monitoring them during execution. - **Counterfactual Tests Reveal Where Small Changes Flip Decisions** — [Source]() — Published 2026-08-12 Tag: Counterfactual Testing Asking what input change would produce a different output shows whether a decision boundary is meaningful or fragile. In clinical studies, counterfactual explanations reduced over-reliance on wrong AI outputs by 21%. The tradeoff: added cognitive load can impair decisions when boundaries are complex. - **Escalation Triggers Define When Humans Must Step In** — [Source]() — Published 2026-08-12 Tag: Escalation Triggers Pre-specified criteria halt autonomous execution and route to a human. The design should weigh reversibility, not just confidence: a high-confidence irreversible action needs a trigger more than a low-confidence trivial one. Triggers that fire too often create the same fatigue they exist to prevent. ### Past Articles - [Web Bot Auth Signs the Sender, Not the Errand]() — By [Sable Whitford]() · Published 2026-08-05 · Format: Protocol Boundaries For years my entire toolkit for identifying a crawler was a user-agent string, an IP range, and a reverse DNS record that any motivated adult could dress up to look wholesome. Web Bot Auth swaps that guess for a cryptographic check, and a narrower one than most coverage suggests. It tells you a key holder sent something to this host inside a time window. Who the agent is working for, and why, are ruled out of scope on purpose. Here's what actually gets signed, and what comes back to your desk. - [Verified, Then Refused]() — By [Sevda Polat]() · Published 2026-08-05 · Format: Purpose Over Detection Site operators are writing robots.txt rules against crawlers that were switched off years ago, and against crawlers nobody has ever identified. You could read that as carelessness, in which case there is nothing to learn from it. Read it as improvisation and it tells you a great deal: people building a classification system that no one has issued them, out of the only material lying around, which is whatever name a client announces about itself. - [Somebody Has to Hold the Basket]() — By [Sevda Polat]() · Published 2026-07-29 · Format: Where State Lives Now A basket, a second item, a card that only the person at the keyboard can produce, and a refund that may take a day to land. Four calls, one story. - [The Exponent Borrows Two Things Nobody Has Measured]() — By [Sevda Polat]() · Published 2026-07-29 · Format: Reliability As Shape Every agent deployment estimate runs the same arithmetic. Per-step success rate, raised to the number of steps, watch the floor open. Directionally right, and nobody should be argued out of it. But it borrows two things the evidence hasn't supplied: one per-step rate to raise, and steps that fail independently of each other. Five runs of the same task, same agent, same instruction, different bill. Horizon is doing most of the damage in the reliability results, and it is the variable those studies measure least directly. ## Other sections in Issue 42 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()