CURRENT | Foundations

Field Guide

Quality Degradation Curves

By Rina Takahashi— December 31, 2025

Feature image for article: Quality Degradation Curves

Your hotel pricing extraction ran at 98% success for six months. Then it didn't. Quality rarely breaks all at once—it erodes. Operating web agents across thousands of sites, you start seeing patterns in how that erosion happens. Some automations cliff overnight when sites change. Others drift downward so gradually you don't notice until you're deep in the hole. The shape of degradation reveals infrastructure capability better than any demo. Understanding these curves matters before you're locked into an approach.

Field Guide

Quality Degradation Curves

By Rina Takahashi— December 31, 2025

Your hotel pricing extraction ran at 98% success for six months. Then it didn't. Quality rarely breaks all at once—it erodes. Operating web agents across thousands of sites, you start seeing patterns in how that erosion happens. Some automations cliff overnight when sites change. Others drift downward so gradually you don't notice until you're deep in the hole. The shape of degradation reveals infrastructure capability better than any demo. Understanding these curves matters before you're locked into an approach.

Tools & Techniques

Tools in Context

When Developers Need to See What's Actually There—The Cookies.txt Browser Extension

Three hours into debugging, you discover the session expired mid-workflow. Figuring that out meant adding logging, rerunning scripts, parsing DevTools. What you needed was simpler: just look at the cookies. Development teams install the Cookies.txt browser extension for this moment. Click the icon, export, open the file. Authentication tokens, expiration timestamps, domain coverage—all visible. No parsing, no code. When iteration speed matters more than automation, seeing what's actually there beats scripting it.

Tools in Context

When Teams Need Sessions That Survive Restarts—Selenium's Cookie API

Test suites that authenticate fresh for every run waste minutes multiplied across hundreds of cases. Automation workflows that break when tokens expire mid-execution lose reliability at scale. Production web agents can't rely on manual cookie extraction—sessions must persist across thousands of runs without human intervention. Teams building at this scale implement Selenium's Cookie API, where code saves authentication state to files and reloads it automatically. When workflows run in CI/CD pipelines, programmatic session management stops being optional.

Tools & Techniques

Tools in Context

When Developers Need to See What's Actually There—The Cookies.txt Browser Extension

Three hours into debugging, you discover the session expired mid-workflow. Figuring that out meant adding logging, rerunning scripts, parsing DevTools. What you needed was simpler: just look at the cookies. Development teams install the Cookies.txt browser extension for this moment. Click the icon, export, open the file. Authentication tokens, expiration timestamps, domain coverage—all visible. No parsing, no code. When iteration speed matters more than automation, seeing what's actually there beats scripting it.

Tools in Context

When Teams Need Sessions That Survive Restarts—Selenium's Cookie API

Test suites that authenticate fresh for every run waste minutes multiplied across hundreds of cases. Automation workflows that break when tokens expire mid-execution lose reliability at scale. Production web agents can't rely on manual cookie extraction—sessions must persist across thousands of runs without human intervention. Teams building at this scale implement Selenium's Cookie API, where code saves authentication state to files and reloads it automatically. When workflows run in CI/CD pipelines, programmatic session management stops being optional.

Tools & Techniques

Tools in Context

When Developers Need to See What's Actually There—The Cookies.txt Browser Extension

Three hours into debugging, you discover the session expired mid-workflow. Figuring that out meant adding logging, rerunning scripts, parsing DevTools. What you needed was simpler: just look at the cookies. Development teams install the Cookies.txt browser extension for this moment. Click the icon, export, open the file. Authentication tokens, expiration timestamps, domain coverage—all visible. No parsing, no code. When iteration speed matters more than automation, seeing what's actually there beats scripting it.

Tools in Context

When Teams Need Sessions That Survive Restarts—Selenium's Cookie API

Test suites that authenticate fresh for every run waste minutes multiplied across hundreds of cases. Automation workflows that break when tokens expire mid-execution lose reliability at scale. Production web agents can't rely on manual cookie extraction—sessions must persist across thousands of runs without human intervention. Teams building at this scale implement Selenium's Cookie API, where code saves authentication state to files and reloads it automatically. When workflows run in CI/CD pipelines, programmatic session management stops being optional.

In Dialogue With Complexity

An Interview with The Handshake

In Dialogue With Complexity

An Interview with The Handshake

Pattern Recognition

Agents Keep Running Out of Context Mid-Task

We're watching companies hit the same wall. Agents start complex tasks, run out of context halfway through, leave work half-finished. Anthropic documented this clearly: even frontier models exhaust their context windows mid-project, forcing the next session to reconstruct what happened.

Here's what matters. Companies are designing agents to work like human developers—continuous sessions, maintained context, marathon coding sprints. Agents need different architecture. They operate in discrete sessions with no memory of what came before.

Look at what's actually working: multi-agent systems, the 89% observability adoption rate. Both point to the same lesson. Agents need explicit session boundaries, clear state management, specialized roles. Not attempts to replicate human work patterns.

Pattern Recognition

Agents Keep Running Out of Context Mid-Task

We're watching companies hit the same wall. Agents start complex tasks, run out of context halfway through, leave work half-finished. Anthropic documented this clearly: even frontier models exhaust their context windows mid-project, forcing the next session to reconstruct what happened.

Here's what matters. Companies are designing agents to work like human developers—continuous sessions, maintained context, marathon coding sprints. Agents need different architecture. They operate in discrete sessions with no memory of what came before.

Look at what's actually working: multi-agent systems, the 89% observability adoption rate. Both point to the same lesson. Agents need explicit session boundaries, clear state management, specialized roles. Not attempts to replicate human work patterns.

Session architecture:

Anthropic's solution uses an initializer agent for setup and a coding agent for incremental progress across discrete sessions, not continuous operation.

Reliability math:

With 95% success per step, a 20-step workflow has only 36% completion probability. Production systems need explicit checkpointing to survive.

Testing gaps:

Agents mark features complete without end-to-end validation, requiring explicit prompting to verify work actually functions as intended before moving forward.

Coordination complexity:

Multi-agent architectures emerged partly because discrete, specialized agents with clear handoffs work better than generalists attempting continuous execution across sessions.

Observability requirement:

94% of production deployments implement tracing specifically to diagnose where agents lose context or make incorrect assumptions about prior work.

Questions Worth Asking

You learn these questions after watching something fail in production. After the third time you've tweaked prompts. After discovering an infinite loop through your monthly bill. After realizing the agent works beautifully until it meets your actual systems.

Vendors answer different questions in demos. They show you feature comparison matrices. They walk through happy paths. But practitioners ask themselves: will this still be running six months from now? Or will it become another expensive lesson in the gap between impressive and reliable?

These are the questions that matter when you're evaluating options.

Questions Worth Asking

You learn these questions after watching something fail in production. After the third time you've tweaked prompts. After discovering an infinite loop through your monthly bill. After realizing the agent works beautifully until it meets your actual systems.

Vendors answer different questions in demos. They show you feature comparison matrices. They walk through happy paths. But practitioners ask themselves: will this still be running six months from now? Or will it become another expensive lesson in the gap between impressive and reliable?

These are the questions that matter when you're evaluating options.

Production Reality

What Breaks Between Demo and Month Three?

Demos run on controlled inputs. Production brings unexpected file formats, unsupported languages, creative ways to break workflows. Edge cases become daily occurrences. Users behave unpredictably. The carefully designed happy path meets reality, and you discover what actually breaks at scale.

Operational Burden

Does This Fade Into Infrastructure?

Some systems run quietly in the background. Others demand constant prompt tweaking, model retraining, manual intervention. Maintenance costs run 30-50% of development annually. Will this become invisible infrastructure? Or will it require a dedicated team just to keep running?

Cost Control

Can You Set a Ceiling on Costs?

Agents enter infinite loops with no error states. They appear normal while racking up token charges. Without depth counters, visited-node tracking, explicit budget controls, costs scale unpredictably. You discover problems through monthly bills, not monitoring dashboards.

Integration Reality

Does This Work With Actual Systems?

Enterprise environments have undocumented rate limits, brittle middleware, custom fields, duplicate logic. Agents tested against clean APIs struggle with real infrastructure. Your organization's actual complexity looks nothing like the idealized systems vendors assume exist.

Debugging Capability

Can You See What Went Wrong?

Agents produce correct outputs through incorrect processes. Silent failures where results look right but execution failed. Without full interaction traces, reasoning logs, tool call visibility, you can't debug production issues. You see that something broke, not what the agent was thinking.

Human Oversight

Where Does Automation End and Judgment Begin?

Agents lack judgment about which decisions need human approval. They confidently execute high-stakes actions without understanding business impact. Human-in-the-loop isn't a fallback. Which decisions require oversight? How do you enforce those boundaries?