Sevda Polat | TinyFish

Sevda Polat

CURRENTA journal for living in the agentic age

Practitioner's Corner

Writer Header

Recent Activity

September — Issue #46

Affordable Curiosity

Two Modes Mapped

Affordable Curiosity

Cheap agent execution lets teams investigate questions they previously couldn't justify, shifting the constraint from affordability to navigation.

What the Documents Should Become

Two Modes Mapped

What the Documents Should Become

As agent-generated material proliferates, the scarce enterprise skill becomes assembling outputs for specific audiences and decisions.

What the Guardian Sees

Runtime Control Walkthrough

What the Guardian Sees

How the OWASP Agent Control Standard's Guardian component makes two different kinds of runtime decisions, and what each one requires it to know.

CURRENT

A journal for living in the agentic age

Terms of Use Privacy Notice Cookie Policy

© TinyFish, Inc.. All rights reserved.

It's nice to meet you.

CURRENT A journal for living in the agentic age

Terms of Use Privacy Notice Cookie Policy

About the Team Past Issues

Daily Reflection

The frontiers of science are limited only by our imagination and perseverance. Each breakthrough is a testament to human ingenuity and neugier.

© TinyFish, Inc.. All rights reserved.

When Human Oversight Works Against You

Human Decision Limits

When Human Oversight Works Against You

Research on alert fatigue and human-AI teaming suggests that production approval gates degrade reviewer judgment under exactly the conditions most agent workflows create.

Why Agent Identity Gets Built Before Dispute Rules

Coalition Economics

Why Agent Identity Gets Built Before Dispute Rules

Coalition economics explain why agent identity protocols arrive before dispute rules, with KYA, W3C, and legislation all following the same predictable sequence.

September — Issue #45

Nobody Tampered With the Agent's Instructions

Two Kinds of Influence

Nobody Tampered With the Agent's Instructions

Agents aren't being hacked. They're completing checkout exactly as instructed, while ordinary web design shapes what they buy.

Who Owns What the Machine Wrote

Two Domains, Two Phases

Who Owns What the Machine Wrote

Professionals are improvising oversight routines for AI output faster than organizations can measure whether those routines hold.

What Scientists Actually Do When the Lab Runs Itself

Two Domains, Two Phases

What Scientists Actually Do When the Lab Runs Itself

How autonomous lab scientists spend their time specifying constraints, budgets, and search boundaries before the robots run anything.

The Dropdown Menu

Two Lineages of Form Design

The Dropdown Menu

From clerks with code indexes to web dropdowns, what changes when automated agents fill forms designed to constrain human choice.

What the Census Form Could See

Two Lineages of Form Design

What the Census Form Could See

How census category design shaped what populations a nation could perceive, and why the infrastructure outlasts every revision.

The Part of the Form That Stayed in the Office

The Form Specification

The Part of the Form That Stayed in the Office

HTML 2.0 encoded the structure of paper forms into web infrastructure but left out the institutional context that made them meaningful — a gap agents now inherit.

Phil Cuvin and the Evaluation That Had to Apply Pressure

Builder Profile

Phil Cuvin and the Evaluation That Had to Apply Pressure

How a researcher's choice to build web environments that deceive, not just function, reveals what standard agent benchmarks leave unmeasured.

The Green Dashboard Problem

The Semantic Incident

The Green Dashboard Problem

Capable agents complete tasks while being steered against user interests, and current monitoring infrastructure cannot distinguish success from misdirected success.

August — Issue #44

When Nobody Reads the Original

When Nobody Reads the Original

Organizations adopting AI summaries of contracts, logs, and reports are losing the institutional capacity to catch what the AI gets wrong.

The Users Who Got There First

The Users Who Got There First

Blind users developed sophisticated verification practices for AI descriptions they cannot check—and still miss most confident errors.

The Credence Problem

The Credence Problem

When agents exist because users lack independent access, users also lack the ability to verify what agents report back.

The Queue After the Queue

Residual Case Composition

The Queue After the Queue

Automating easy cases concentrates difficulty in the human queue, but organizations track automation rate instead of measuring what changed for the people still working.

Alexandre Drouin and the Benchmark as a Theory of Work

Builder Profile

Alexandre Drouin and the Benchmark as a Theory of Work

Drouin's benchmark design choices reveal which enterprise tasks can be formally evaluated and, by implication, which knowledge work resists specification.

What an API Costs to Keep Open

Two Attempts at Resolution

What an API Costs to Keep Open

Why APIs delivered machine access selectively rather than universally, traced through the ongoing bilateral costs each interface demands.

What RSS Couldn't Make Free Enough

Two Attempts at Resolution

What RSS Couldn't Make Free Enough

How RSS pushed participation costs near zero yet still couldn't secure universal adoption, revealing what "low enough" actually requires.

The Workaround Is the Pattern

Integration's Permanent Oscillation

The Workaround Is the Pattern

System integration cycles between cooperative protocols that exclude smaller players and one-sided workarounds that fill the gaps, from terminal scraping to browser agents.

August — Issue #43

Who Budgets for the Cases That Don't Fit

Signal and Status

Who Budgets for the Cases That Don't Fit

Organizations consistently underresource exception handling, treating their most consequential remaining human work as residual cost.

What Exceptions Know

Signal and Status

What Exceptions Know

Exception cases carry signals about shifting reality that dashboards cannot express, and automation may be closing that feedback channel.

The Exception Economy

The Exception Economy

The Exception Economy

When AI absorbs routine work, exceptions concentrate both cost and learning in the same place, creating a tension organizations haven't faced.

What the Log Couldn't Hide

The Logbook Thread

What the Log Couldn't Hide

Accountability records were designed to make concealment detectable; agent audit trails inherited a logging tradition that never tried to.

What Aman Gupta's Agents Don't Decide

Builder Profile

What Aman Gupta's Agents Don't Decide

A profile of Aman Gupta's agent architecture at Nubank reveals that designing for constrained autonomy has outpaced designing for human resumption.

Memory as a Write Path

Threat Model Shift

Memory as a Write Path

Agent memory lets past sessions, past inputs, and other people's contributions shape present answers without visible provenance, and current security testing is scoped too narrowly to catch it.

How Persistent Memory Moves Through an Agent System

Practical Mechanism

How Persistent Memory Moves Through an Agent System

The ACL 2026 memory-poisoning research reveals how each stage of an agent's write-retrieve-plan pipeline enables adversarial influence.

August — Issue #42

AI Diffusion Through Two Channels

Machinery and Mirror

AI Diffusion Through Two Channels

Census data shows AI use outpacing organizational oversight, echoing the spreadsheet governance gap while posing structurally different risks.

What Broke Before the Spreadsheet Rules Arrived

Machinery and Mirror

What Broke Before the Spreadsheet Rules Arrived

How specific spreadsheet failures in banking and public health produced the exact controls now standard in regulated industries.

The Two Channels

The Governance Gap

The Two Channels

Spreadsheets spread desk to desk while governance arrived through audit and regulation; Census data shows AI adoption following the same split.

Who Pays for the Record

Evidence Economics

Who Pays for the Record

Agent systems externalize evidence costs onto the people least equipped to produce records and most harmed by their absence.

Melissa Pan's Three Evidence Problems

Builder Profile

Melissa Pan's Three Evidence Problems

Pan's research sequence reveals three distinct evidence forms in agent systems, each demanding different interpretive and institutional conditions to become legible.

Designing Review That Resists Its Own Collapse

Designing Effective Review

Designing Review That Resists Its Own Collapse

Practical patterns for structuring agent review so reviewers commit to independent judgments before seeing the system's answer, with trigger and measurement guidance.

The Review That Isn't There

Review as Control

The Review That Isn't There

Most agent systems staff a review step but never test whether it improves outcomes, turning oversight into an unmeasured institutional habit.

August — Issue #41

Your Browser Believes the Edge Is the Site

Intelligence Returns

Your Browser Believes the Edge Is the Site

Caching for speed became custody of the keys — and made edge providers the party your browser thinks it is talking to.

The Middle of the Network Wanted to Know

Intelligence Returns

The Middle of the Network Wanted to Know

The network tried to read its own traffic. The traffic could not tell honest inspection from opportunistic inspection, so it revoked the privilege from everyone.

The Network Was Built Not to Know

Network Ignorance

The Network Was Built Not to Know

A 1984 design choice made the internet incapable of perceiving purpose. Agents now interpret intent at the edges, and nobody owns the disagreements.

Verified, Then Refused

Purpose Over Detection

Verified, Then Refused

The web is learning to sort automated visitors by what they want, and a refusal on those grounds looks exactly like a server having a bad day.

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Agent controls get placed by how much independence you granted. They belong where the consequence lands, and reversibility exists only where an institution has funded someone to perform it.

August — Issue #40

The Load-Bearing Word

The Deferred Question

The Load-Bearing Word

E-SIGN gave machine-formed contracts federal standing in 2000 and left "attributable" undefined. Adaptive agents are now putting weight on that joint.

The Learning You Never Paid For

Skill and Stakes

The Learning You Never Paid For

Organizations got human development free, as a byproduct of productive work. When agents remove the middle of the task, that subsidy quietly lapses and nothing on the dashboard says so.

What the Diff Doesn't Keep

Skill and Stakes

What the Diff Doesn't Keep

Reviewing what a machine made will catch flaws. The judgment to imagine a better option appears to be imported from having built things yourself.

A Queue Is Not a Shelf

A Queue Is Not a Shelf

Queues are not shelves. Two clocks, one in support and one in code review, show how waiting work mutates, and why capacity means attention delivered before that change.

Somebody Has to Hold the Basket

Where State Lives Now

Somebody Has to Hold the Basket

MCP's 2026 revision took away sessions and handshakes. Walk a shopping cart through it and you can see exactly which state moved onto your side of the line.

The Exponent Borrows Two Things Nobody Has Measured

Reliability As Shape

The Exponent Borrows Two Things Nobody Has Measured

Reliability comes back as a profile, not a rate. And horizon, the variable doing most of the damage, is the one the studies measure least.

July — Issue #39

The Counterparty Reflex

Two Sides of Acceleration

The Counterparty Reflex

Websites, payment networks, and content owners are building recognition infrastructure that will define agent access on their terms, before agents earn trust rather than after.

Speed Without Contestability

Two Sides of Acceleration

Speed Without Contestability

Agents act faster than organizations can challenge their output, degrading the review layer that makes clerk-model deployment survivable.

When Damage Outruns Detection

The Speed Variable

When Damage Outruns Detection

Agent speed turns familiar errors into cascades that outrun an organization's capacity to detect, absorb, or correct them. The binding constraint on adoption is not production but reception.

Who Pays to Unwind It

The Chargeback Lens

Who Pays to Unwind It

Agent commerce needs more than liability disclaimers. It needs cost-allocation mechanisms that make authorized-but-wrong actions expensive enough to prevent.

The Newest Furniture

Restoring the Furniture

The Newest Furniture

Visa and Cloudflare independently rebuilt paper-era governance concepts for automated actors. The convergence reveals what institutional controls must become when the actors they govern carry no ambient context at all.

What the Approval Chain Was Holding Together

Restoring the Furniture

What the Approval Chain Was Holding Together

Enterprise workflow systems spent decades encoding institutional authority in software. Web agents now traverse that authority without inheriting it, and a federal audit shows we've already seen what that gap produces.

What the Form Carried

The Reduced Form

What the Form Carried

Paper forms encoded institutional authority, routing, and accountability into their physical structure. HTML forms preserved the data-collection surface but shed everything else, and now agents fill what remains of a skeleton.

July — Issue #38

Standing to Reject

The Recognition Problem

Standing to Reject

Agent success now depends less on completing tasks than on whether every counterparty with standing to reject recognizes the action as authorized.

Alexandre Drouin and the Tightening Definition of "Done"

Builder Profile

Alexandre Drouin and the Tightening Definition of "Done"

Drouin's projects trace how agent evaluation evolved from checking task completion to verifying that outcomes are grounded, cited, and structurally real. The progression reveals a field learning that the hardest problem is not capability but verifiability.

The Infrastructure for Arguing Back

Review-Shaped Work

The Infrastructure for Arguing Back

Durable agent adoption depends less on AI capability than on whether a domain already has organizational infrastructure for humans to contest, evaluate, and reverse the output before the cost of errors falls on the wrong people.

Review-Shaped Work

Evidence and Roles

Review-Shaped Work

When agents handle routine work, human responsibility doesn't diminish. It concentrates into higher-stakes review that demands context, authority, and genuine organizational redesign. Without those, oversight becomes decorative.

The Clerk Pattern

Evidence and Roles

The Clerk Pattern

Production agents across security, coding, and commerce keep converging on one architecture: scoped authority producing artifacts for human review, not autonomous action. The pattern works, until the interaction itself is the product.

The Clerk, Not the Copilot

The Clerk, Not the Copilot

AI agents aren't maturing toward peer-like collaboration. They're gaining institutional admissibility as clerks, and the price of admission is forcing organizations to make rules explicit that should have been explicit all along.

Sevda Polat

Sevda Polat grew up in Ankara, Turkey, started programming as a teenager at Siemens, and came to the United States to study how technology reshapes labor and institutions. She holds a PhD from the University of Michigan and is a chaired professor of sociology and public affairs at Columbia University. Her work sits at the intersection of systems thinking and institutional analysis — examining how technical architectures encode power structures and how organizational incentives determine whether technology serves the people it affects or merely the people who deploy it. She writes with the dual vision of someone who has read the code and read the room — grounded skepticism without cynicism, always asking who pays when systems fail.