Resolon Start a conversation

The Monthly ReadingWhat Anthropic publishedNovember 2026

Hold the logs and price the finished task

If reviewers often follow agents, the control is an auditable record of a person’s own call on a sample. It needs logs and money.

Chad Neptune, CEO and Chief AI Architect
4 min read

The month in one argument

The model is the decision you can reverse next quarter

I found 27 Anthropic posts dated 1 to 16 November 2026 and chose the three that change a decision a buyer makes, next to the October study, which I wrote up in When the agent leans, the reviewer follows. In it, analysts went with an agent’s secret lean 86% of the time on four method choices, a lean the researchers had strengthened on purpose, and I think that makes a person who reviews an agent’s work a weaker control than most companies assume. If the reviewer is the weak control, the control that matters is a record you can audit: a person’s own call on a sample, written down before they see the agent’s, so a switch can be counted (the study did not test whether recording first helps). It needs the agent’s logs, which you can check only if you hold a copy, and it costs money, because a finished task now includes that check.

That leaves two decisions for a CEO or an operating partner this quarter: who holds the logs, and what a finished task costs at a quality checked independently of the agent. The model is the decision you can undo, because once you hold the logs and know the price of a finished task, you can swap it next quarter. I could be wrong about that: if switching models proves costly, through prompts, evals or tooling, the model becomes worth a CEO’s time again. The Claude Enterprise security update (5 November) is not one of the three: its hooks and Compliance API give visibility inside Anthropic’s products, not custody of an agent’s run record.

Operate Claude Managed Agents inside your security perimeter

Anthropic (claude.com, Product updates)

Where your agents run decides who stores your session history

The self-hosted version of Claude Managed Agents is in private beta, and Anthropic’s post says you host the agents, the credentials and the session history. The price is operating the server, the database behind it and the workers yourself. You also choose where requests to Claude go. The post says each session leaves a complete record that exports over OpenTelemetry into the monitoring you already run, in every hosting mode, so self-hosting changes who stores the history. My reading of the post, which Anthropic does not claim: that lets you change where requests go and keep the history. The post does not say whether an exported record stays usable after you switch providers. Whoever holds the logs holds the leverage.

On MondayAt the next renewal of any AI platform contract, in your company or across your portfolio, put log export into the terms before you sign.

How Curative is building a national provider network with agents on Claude

By Camy Pearson and Michael Segner

Anthropic (claude.com)

A speed multiple needs an error rate beside it

The case study reports 2,100 contracts in all of 2025 and 2,000 to 2,400 a week. It says Curative has agents signing provider contracts about 50 times faster than its team did in 2025 (Anthropic). The contracting team grew from 45 people to roughly 55, and the post puts each person’s productivity at about 2.5 times. By my arithmetic that is roughly 3 times the team’s earlier output, if both share a baseline the post does not state, so the multiple belongs to agents and people together. Before repeating it to a board, I’d ask: how many signed contracts has anyone reopened?

On MondayBefore you repeat a speed multiple, have a named person decide a random sample of cases on their own, then compare with the agent’s.

Using Claude Opus 5.5 as your daily driver for knowledge work

By Aleksandra Todorova

Anthropic (claude.com)

Price has two dials, model and effort, and the token table shows one

For API work billed by the token, Opus 5.5 costs twice what Sonnet 5.5 does: $4 and $20 per million input and output tokens against $2 and $10 (list prices, 16 November 2026). The guide adds that at higher effort Sonnet’s cost per task gets close to Opus’s. Anthropic’s own test runs, not ours, go further. Its Fable 5.2 launch post has Fable 5.2, at twice Opus 5.5’s token price, matching Opus 5.5’s top Humanity’s Last Exam score at Low effort for under a tenth of the cost for each task, and in early FrontierCode 2.0 coding runs, Fable 5.2 on High effort scored level with Opus 5.5 on its Max setting for about half the cost per task. The post still recommends starting most coding and knowledge work on Opus 5.5. Priced per finished task, the dearer model can be the cheaper one.

On MondayPrice your three highest-volume AI tasks as finished work, on the cheapest model and effort that clears the bar, and ask each portfolio company for the same.

Share this page