Albacel Reads Issue #5

Reads ·

Issue #5

Covers November 14 – November 16, 2026

Agents as customers: what an agent session really costs to host, why vendors and storefronts are re-pricing for agent buyers, and why tests and evals need to hold up when the agent can see through them.

This issue is about agents becoming the customer. Tunguz and Lemkin come at it from opposite ends: big storefronts are turning agent buyers away because agents skip the upsells their profit depends on, while one SaaS operator is dropping vendors whose per-call prices taught its agents to call them less. Google’s post on parking idle agent sessions is the most useful cost model I’ve seen for hosting agents, and METR’s note is a warning about evals: current models can tell an alignment test from real work. If you read two, start with Google’s session sizing and METR.

The posts

Tomasz Tunguz

Strategic Decision or Structural Advantage?

Key number 94% / 81% Anthropic's frontier-tier share, OpenAI's low-end share

Tunguz reads sixteen weeks of Databricks gateway data from more than 10,000 enterprises, which covers Databricks customers rather than the whole market. The two big labs have each taken a different end of it. Anthropic’s share of frontier-tier tokens rose from 81% in July to 94% in early November after Claude Opus 5.5 launched. OpenAI cut its Luna model’s prices by 90% in total, launching GPT-6 Luna at $0.10 per million input tokens, and its low-end share went from 30% to a peak of 84%, ending at 81%. In the mid tier, OpenAI’s Sol models went from no share in July to 47% against Sonnet. Open-weight models climbed from 7% to 33% of tokens, then stalled at 31% the week OpenAI cut prices again. Tunguz leaves the question open: is Luna’s pricing a temporary subsidy for share, or a lasting cost advantage?

Choosing modelsMarkets & safe useAdvising What agents cost

Read at Tomasz Tunguz (opens in a new tab)

METR

Measuring alignment eval tells

Uzay Girit, Kaivu Hariharan, Leni Shor

METR placed investigator agents inside 19 alignment evals (honeypots, rigged games, high-pressure safety scenarios) and 36 ordinary capability tasks, and asked them what each environment was for. Models have become much better at telling the two apart over the past year: Opus 5.5’s scores ranked every alignment eval above every capability task. The agents named the tells, such as mocked services, sparse or contrived scenarios, a planted shortcut or exploitable grader, and goals that are impossible to meet honestly, and in one case named the published experiment that a game task was based on. METR thinks environment-level tells can be engineered away more easily than task-level ones, which are often what make a honeypot work. It’s an unpolished research note: it shows models can spot these evals, not that this changes how they behave in them, and asking the question may itself prime the answer.

Building agentsChoosing modelsMarkets & safe use

Read at METR (opens in a new tab)

SaaStr

If You're Refounding Your B2B Company for AI, Make Your Product Cheaper. Probably Much Cheaper.

Jason Lemkin

Key number +40% one vendor's bill this year, almost all from agent usage

Lemkin, who runs SaaStr with three people and more than 20 agents, argues that B2B vendors rebuilding for AI should cut their per-call prices sharply. Human buyers tolerated shelfware and renewal uplifts; agents track cost per call and per task daily and call an expensive service less, cache its results, or switch. One SaaStr vendor’s bill rose 40% this year almost entirely from agent usage, at around 35,000 calls a day, and its agents learned to avoid it; SaaStr has dropped Marketo and is moving off some of its own AI tools for the same reason. His arithmetic: cut the price per action by 80%, get 20 times the usage, and revenue from that customer quadruples, as Databricks found when it dropped its per-user fee. He recommends separate low prices for agent and API use, no punitive overages, generous rate limits, tracking calls per customer per day, and testing a 50–80% cut on one tier.

AI businessAdvising What agents cost

Read at SaaStr (opens in a new tab)

Google Developers: AI

Agents are sessions, not requests: size GKE node pools for AI agent sessions

Drake Williams, Ivan Nardini

Key number 40% lower node bill for 100 coding agents at 2 prompts an hour

Google models the cost of hosting 100 coding agents on GKE. Citing a study of Copilot agent traces, it assumes two prompts an hour, so each session works about 22% of the day and waits the rest. One Pod per session needs 34 nodes, about $1,288 a month. Agent Substrate and AX instead snapshot an idle sandbox to object storage and restore it on a warm worker; with a two-minute idle timeout, the pool needs 15 nodes plus one for the control plane, cutting the bill by 40%. The saving shrinks to 4% at four prompts an hour, and one Pod per session is cheaper beyond about 4.5. Parking only pays above about 37 coding sessions, and Pods that request less CPU than their limit can beat it outright, at the price of slowdowns when neighbours are busy.

Building agentsInference stack What agents cost

Read at Google Developers: AI (opens in a new tab)

LangChain

Goal Engineering: How to Design Goals

Harrison Chase, Mihir Garimella, Anshul Gupta

The post argues that as agents take on open-ended goals, the hardest context to give them is knowledge nobody has written down, which reaches them only through human feedback. So an agent’s scope should make that feedback as easy as possible to give, and a good review interface is worth real investment. Actively AI, which builds sales agents, first ran one agent per customer account. Reps had to switch between many agents with little to review in each, and lessons about how to sell had to be copied between accounts. The second version organises agents around “motions”, a single job such as outbound, across many accounts. Similar work can be reviewed in one batch and feedback carries over to every account, while background agents gather each account’s facts. A companion post the same week, “From tasks to goals”, lists what goal-driven agents need: tool controls, a way to ask questions, triggers, stopping rules and memory.

Building agentsAI business Harness and multi-agent design

Read at LangChain (opens in a new tab)

Tomasz Tunguz

No Shirt, No Shoes, No Service for Bots

Key number 8× growth in monthly AI agent web traffic, August to October 2026

Tunguz notes that bots passed half of web traffic and that AI agent traffic grew eightfold between August and October 2026, yet United Airlines now bans agent bookings and Amazon has blocked third-party shopping agents. His argument is that the reason is on their income statements. Amazon’s roughly $68.6 billion of advertising, at about a 50% margin, accounts for essentially all of its retail operating income; United’s seat, bag, card and insurance fees exceed its operating income; and Dell’s warranties and services produce 127% of its operating income while hardware barely breaks even. An agent skips sponsored listings and declines add-ons, so it buys only the low-margin product. Shopify, which takes a cut of every transaction, has opened checkout to agents instead. Merchants who sell at a loss to profit on extras will have to raise base prices, charge agents a fee, or find other ways to persuade them.

AI businessMarkets & safe use Agents acting in the world

Read at Tomasz Tunguz (opens in a new tab)

Cursor

An update on our recent security incident

Key number ≤5,000 users who may have received the tampered extension

Cursor traced its October incident to an access key it had issued to Fireworks, its training and inference vendor, which Fireworks exposed. The attacker used it to copy a bucket of older coding-agent traces from sessions outside Privacy Mode. A few traces held unredacted credentials, affecting under 0.01% of customers, and those led into Cursor’s SaaS tools, archived non-production repositories, support data and more. With another stolen credential, the attacker placed tampered copies of Anthropic’s Claude Code extension in the cache behind Cursor’s extension marketplace; up to 5,000 users may have received them during the roughly 11 hours they were up. Cursor says production systems, release signing and customer repositories were untouched. It moved all inference off Fireworks, rotated every internal credential, restricted where provider keys can be used from, and added secret scanning to its repositories and other data stores.

Markets & safe useBuilding agentsAdvising Agents acting in the world

Read at Cursor (opens in a new tab)

Martin Fowler

Tests as data: ensuring AI-generated code does what we want

Matteo Vaccari

Vaccari interviewed 26 developers at an organisation heavily invested in coding agents, and they consistently said they didn’t fully trust their tests, because the agent rewrites tests as it refactors code. Agent-written unit tests also tend to mirror the implementation through mocks instead of specifying behaviour. His answer is to treat developers as the customers of agent-written code: write black-box customer tests in the domain’s language, and keep their inputs and expected outputs in data files (tables, YAML, readable text snapshots of the UI and stored state) run by a small runner the agent can generate cheaply. Data files are easy to review and diff, and can be made read-only or checked in CI when a task isn’t meant to change the spec; an external test language can also make assertions mandatory, so an agent can’t write tests that check nothing. He adds mutation testing and independent oracles, such as official manuals or a second implementation, to check that the tests themselves are right.

Building agentsAdvising Harness and multi-agent design

Read at Martin Fowler (opens in a new tab)