ClawCamp · AI Infra Summit · Aug 2026

01 / 07

Only 18¢ of your AI dollar reaches the product.

The model isn’t your bottleneck. Your routing is. So is your context window.

Paper 1 · Reliable AI Systems

01 · The Number

02 / 07

For every dollar of tokens, 82 cents never reaches production.

Self-reported survey across 2,444 companies.

44¢

fixing bugs the AI created

27¢

rewriting AI-generated code

11¢

review and merge delays

18¢

reaches production

01 · The Number

03 / 07

That number is directionally right and numerically dubious.

Better you hear the objection from me than find it afterwards.

Why to doubt it

The vendor who published it sells the fix

Self-reported, likely including free tiers

Maintenance was already 75 to 80% of effort

Before AI existed. So 82% is close to the historical baseline, renamed.

What is not in dispute

1.7× more issues in AI-written changes

CodeRabbit

43% still needs manual debugging after passing quality checks

Lightrun, 2026

Generation scales. Human review does not.

The waste did not change. The ratio did.

AI let you generate defects faster than anyone can review them. The bottleneck was never cost per token. It is review throughput.

02 · The Workaround

04 / 07

So how are you getting around it right now?

Hands up if any of these is yours.

Two accounts. Maybe three.

Because one runs out before lunch.

Three tools that do the same job

So there is always one with quota left.

Waiting out the rate limit

Laptop open. Nothing shipping.

Starting over, re-explaining everything

Because the session “got dumb.”

None of these are workflow preferences. Every one is you buying more capacity to absorb the waste, instead of removing it.

03 · The Fix

05 / 07

Stop paying judgment prices for clerical work.

Four moves. You can start all four on Monday.

01

Make the prompt earn its tokens

A vague prompt buys a wrong answer at full price, then you pay again to fix it. Co-Dialectic does this on every turn. Free and open source.

02

Never let a model grade its own homework

Same family, same training, same blind spot. It will confidently approve its own mistake. Judge across families.

03

Right-size the model to the task

Whale for reasoning. Fish for the mundane. Deploys, file reads and git operations do not need a frontier model.

04

Right model, right job. They are not interchangeable.

Fable 5

Reasoning

Claude

UI and design

Codex

Follows instructions

Gemini Flash

Chrome, 50× cheaper

04 · The Other Leak

06 / 07

A bigger context window makes your agent dumber.

Not slower. Not just pricier. Worse at the task.

What the research found

Bury the answer in the middle and the model loses it

Liu et al., Lost in the Middle, TACL 2024

Irrelevant context does not dilute. It competes.

Semantic similarity drives the decay more than length does.

Models scored better on shuffled context than on coherent context

Across 18 models. Coherent documents share vocabulary, which makes them better distractors.

What to do instead

Sharp focus

Keep the working context deliberately small.

Long-term memory

Everything else lives in an index outside the window.

Just-in-time injection

Only the relevant slice, only when the task needs it.

Compress by relevance, not volume

A reset becomes a checkpoint instead of a loss.

“It fits” is not the same as “it helps.” The more relevant-looking the material you stuff in, the more damage it does.

05 · Or Automate All Four

07 / 07

xTeamOS is the agent operating system for your company.

Routing, gates and cross-family judging on every change. With your own keys.

100% private

Your work never leaves your control

Portable

No lock-in to one vendor or model

Federatable

Across people, teams and companies

Development

Cost routing, gates, cross-family review

Ideation

Sharper thinking before code

Branding

Your company voice

Sales

Pipeline and relationships

Beta

Open now, and pricing is not set.
I am here to find out what this is worth to the teams who would actually run it. Come tell me.

Co-Dialectic stays free and open source.

Try xTeamOS

xteamos.exponentialos.io

Read the whole argument

thewhyman.blog