Skip to main content
Discuss a project日本語
HAKOMIRO

Daily AI use edition

Type
Launch catch-up (a look back over the covered period, not events of the day)
Edition date
Nov 1, 2026 (Japan time) | Version 1
Period covered
2026-10-26 to 2026-11-01
Checked through
2026-11-01 18:43 JST
Last checked
2026-11-01 18:56 JST

Today’s key points

Launch look-back published

This is the catch-up edition for the founding issue. We read six primary sources published October 26 – November 1, 2026, and included the parts that bear on judgments at work. They are not events of November 1 itself.

What was searched and its limits: This covers six primary sources published October 26 – November 1, 2026, read through to the body text. There are four business and two technical items (about 67:33), and the guide for the last seven days is roughly 7:3. Subjects are a UK local authority, company surveys in Spain and Mexico, a US company's worldwide office workers, and Japanese-language technical experiment articles. The location of the authors of the two technical articles is unconfirmed. The source languages checked were Japanese, English and Spanish; it is not a survey covering every region and industry in the world.

Balance of coverage (last 7 days, independent new items): work and society 4 / building and development 2. The target is roughly 7:3; we do not add items just to match it.

New | Work and society

UK rough-sleeper reporting service StreetLink: operators report fewer "person not found" reports after AI guidance at input

What it means for work (editorial judgment)
This practical case suggests that if you change the design so that, at the point of input, AI asks follow-up questions about what is missing, instead of people fixing entries afterwards, wasted visits in the field may fall. It can be used to review situations where input quality causes rework downstream in your own inquiry, application or reporting intake. However, the figures are the operators' self-report and are not an effect that necessarily applies to your organization.
What changed
The UK Ministry of Housing, Communities and Local Government (MHCLG) and the service provider Home Connections trialed AI input guidance in 10 local authorities for StreetLink, where the public reports rough sleepers in England and Wales. When a description of a location is vague, it prompts for more, such as landmarks and ways of access. According to the report, failed visits due to missing information went from 36% to 18%, the share of reports after which the person was found went from 16% to 23%, and descriptions of the location grew by 45% and of the person by 54%. About 89% of users said it "helped them write in more detail", and the report says a full rollout has been carried out.
Who it concerns
People in government or companies who run reception, counters, call centers or application forms, and the departments that respond on site to reports or applications.
What to check or decide today
First count the cases where missing input at your intake causes rechecks or wasted visits, and their reasons. Try the idea of AI asking follow-ups in the input field first in a small scope, in a form that leaves judgment with people. Do not use this article's figures as expectations for your own organization.
Limits
This is the operators' self-report, and we have not confirmed third-party verification. The article does not give the comparison period, the denominators such as the number of reports, or whether there was a control group. We cannot say that the improvement in the "found" share is due to the AI guidance alone. The figures do not show the effect on protecting or helping rough sleepers. It is stated that the design leaves decisions to staff.
Result
Failed visits due to missing information 36%→18%; share of reports after which the person was found 16%→23% (unit: Percentage (%) | period: No comparison period given (trial in 10 local authorities) | basis: Provider's claim)
Cost (partial only, not a total)
Unknown: The primary source gives no introduction, operating or checking costs, so the amount is unknown. It cannot be read as free or zero.
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Not confirmed in the source
  • Data preparation: Not confirmed in the source
Stage of the effort: Adoption
The primary source says a full rollout was carried out after the trial in 10 local authorities.
  • What was expected: Reduce failed visits with guidance at input
  • What was observed: Failed visits from 36% to 18% (operator's report)
  • What to check next: Whether the same level continues after the full rollout, and whether anyone other than the operators has verified it.

A small trial to try (the editors’ proposal, not a result we ran)

Goal: Before using AI, count the rework that missing input at intake creates downstream

Prerequisites

  • Staff who can read intake records
  • Intake records of the last one to two months

Steps

  1. Count the cases that became rework, split into "missing input" and "other"
  2. Write out the top three reasons for missing input
  3. For those reasons, draft on paper one follow-up prompt to add to the input field (do not introduce AI)

What success looks like

  • The numbers show that missing input accounts for a large share of all rework
  • The draft follow-up prompts pass checking by staff in the field

When to stop

  • The reasons cannot be classified from the records
  • If missing input is only a part of the rework, do not proceed to AI guidance and investigate other causes

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
UK (government and contractor)
Region concerned:
10 local authorities in England and Wales
Source language and what was read:
English / Read the full English original blog text (no translation used)
HAKOMIRO’s check — confirmed conditions for Japan:
For Japanese readers, only an outline of the idea (asking follow-ups about missing information at input) could be confirmed.
HAKOMIRO’s check — not confirmed for Japan:
The article does not state the name, cost, Japanese-language support or personal-data handling conditions of the AI service used, so these are unconfirmed.
Technical details and handoff

This is input guidance that asks follow-up questions about landmarks and ways of access when a description of a location is vague. The primary source does not state the AI service or model used, or how personal information is handled.

Objective:
Organize the requirements for a mechanism that returns additional questions when input is insufficient
Inputs:
List of intake form fields; Past examples of missing input (excluding personal information)
Expected output:
Drafts of follow-up prompts for each kind of gap, and the conditions under which judgment stays with people
Constraints:
Staff judgment stays in place; First check where personal information is sent and the storage conditions; The service used cannot be known from the primary source
How to verify:
Compare the completeness of input with and without follow-ups; Spot-check for errors in content that the follow-ups led people to enter

New | Work and society

Spanish companies' AI use: 33.6% of companies with 10+ employees in 2025, with a large gap by size (Cotec report)

What it means for work (editorial judgment)
The story that "more companies are using AI", regardless of your size, looks very different by size. Looking at the gap between large and micro companies and the reasons given by non-users gives a way to check, starting from "people and knowledge", why your own adoption is lagging. However, these are Spanish statistics and not figures that apply directly to Japanese companies.
What changed
Cotec, the Spanish innovation foundation, published a report on October 26, 2026. It analyzed public statistics (INE) for full-year 2025 and Q1 2026: the share of companies with 10+ employees using AI rose from 21.1% in 2024 to 33.6% in 2025 (+12.5 points). In 2025 it was 74.4% for large companies and 21.4% for micro companies, a gap of 53.0 points. Spanish companies using AI are estimated at about 726,000, and micro companies make up 92.6% of the users. The biggest barrier cited by non-users is a lack of in-house knowledge, named by about 80%.
Who it concerns
Owners of small and medium-sized companies, and AI adoption leads at companies with trading partners of different sizes.
What to check or decide today
Check why your company is "not using" it by starting not with cost or technology but with "do we have people and knowledge in-house who can make the judgment". For comparison, use only figures for your own size, and do not make the overall average a target.
Limits
These figures are companies' answers to a statistical survey and do not measure AI's outcomes or costs. AI counts companies using any of eight kinds of technology, so it is not a figure for generative AI alone. Regional and industry analysis is limited to companies with 10+ employees. We have not checked how well this applies to Japanese companies.
Result
AI use by companies with 10+ employees 21.1%→33.6% (+12.5 points). In 2025: large companies 74.4%, micro companies 21.4% (unit: Percentage (%) and percentage points | period: 2024 and 2025 (analysis includes Q1 2026 data) | basis: Survey response (self-reported))
Cost (partial only, not a total)
Unknown: The announcement and published text we checked this time do not give actual amounts such as introduction or operating costs. The cost items in the questionnaire and full report are unconfirmed.
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Not confirmed in the source
  • Data preparation: Not confirmed in the source
Stage of the effort: Unknown
These are aggregate statistics and do not show the stage of any individual effort.
  • What was expected: Not stated
  • What was observed: Use among companies with 10+ employees went from 21.1% to 33.6% (companies' answers)
  • Why unknown: These are overall statistics, not a specific initiative, and no expected level is stated.
  • What to check next: Whether use is established at smaller companies, and what happened as a result of continued use.

A small trial to try (the editors’ proposal, not a result we ran)

Goal: Check whether the reason your company is not using AI lies in in-house knowledge or judgment

Prerequisites

  • Head of the responsible department
  • Three candidate AI uses you would like to try at work

Steps

  1. For each candidate, write out "who decides and who checks"
  2. Mark the candidates for which there is no one who can decide
  3. Make the marked candidates the target of ways to build knowledge (training, outside consultation)

What success looks like

  • A decider and a checker are set for each candidate
  • The candidates for which lack of knowledge is the reason can be identified

When to stop

  • Stop considering adoption for candidates where no decider can be placed
  • Do not move on to cost or contract discussions until this check is finished

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
Spain (innovation foundation Cotec)
Region concerned:
Spanish companies
Source language and what was read:
Spanish / Read the announcement page and report page text (original read, no machine translation used)
HAKOMIRO’s check — confirmed conditions for Japan:
For Japanese readers, only an outline of the tendency that adoption differs by size could be confirmed.
HAKOMIRO’s check — not confirmed for Japan:
Comparable figures for Japanese companies, differences from Japan in how the survey population is defined, and the full tables of the report are unconfirmed.
Technical details and handoff

Using survey data from the public statistics INE, companies using any of eight kinds of AI technology are counted as users. Regional and industry analysis covers only companies with 10+ employees.

Objective:
Choose the statistical category matching your size and organize it so it can be compared
Inputs:
The report's tables by size; Your company's employee size
Expected output:
Usage rate for your own size category and a note on definitions
Constraints:
The definition of AI technology is broad; It is not a figure for generative AI alone; The definition may differ from Japanese statistics
How to verify:
Check that the figures you cite match the report text; State the definition and the scope of companies together

New | Work and society

KPMG survey: 90% of large Mexican companies say AI delivers value, but only 14% have a mechanism to compare it with investment

What it means for work (editorial judgment)
There is a large gap between the self-assessment that "AI is creating value" and the share that has a mechanism to compare that value with the amount invested. When reporting AI outcomes in your own company, it gives a way to check whether outcome figures and costs are set side by side on the same basis. However, these are self-reports by respondents at large companies and do not measure actual outcomes.
What changed
In October 2026, KPMG Mexico announced the Q3 2026 results of the "Global AI Pulse", a survey of more than 2,100 executives in 20 countries (conducted July–August 2026, covering companies with annual revenue above US$50 million). In Mexico, 37% of companies (34% globally) had deployed AI agents on a large scale. 90% (89% globally) said they were getting measurable value from AI, but only 14% (12% globally) have a consistent mechanism to compare value with investment. In Mexico, 41% plan to invest up to US$499 million over the next 12 months.
Who it concerns
People who report the effect of AI investment to executives, and staff in corporate planning and finance at large companies.
What to check or decide today
Check whether the outcome reports for your AI initiatives list costs (introduction, checking and fixing, operation, data preparation) on the same basis. For initiatives where only outcome figures come first, set out the costs side by side before evaluating.
Limits
This is a survey based on respondents' self-reports, and neither value nor investment amounts were measured by a third party. The investment amounts are plans, not actual spending. The press release does not give the number of respondents in Mexico or the denominators for each question. It covers companies with annual revenue above US$50 million and does not apply to small businesses. We have not checked Japanese companies.
Result
Mexico: deployed AI agents on a large scale 37% (global 34%); getting measurable value 90%; have a mechanism to compare value with investment 14% (unit: Percentage (%) | period: Survey of July–August 2026 | basis: Survey response (self-reported))
Cost (partial only, not a total)
Unknown: The announcement text we checked this time has no actual cost or total cost figures. The cost items in the questionnaire and full report are unconfirmed. Planned investment over the next 12 months is not treated as actual cost.
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Not confirmed in the source
  • Data preparation: Not confirmed in the source
Stage of the effort: Unknown
The survey aggregates answers from large companies and does not show the stage of any specific initiative.
  • What was expected: Investment plans for the next 12 months (41% of companies have plans of up to US$499 million)
  • What was observed: 14% have a consistent mechanism to compare value with investment
  • Why unknown: These are statistics for the survey as a whole, and the stage of each initiative is not given.
  • What to check next: Were the planned investments carried out, and were their outcomes evaluated against the investment?

A small trial to try (the editors’ proposal, not a result we ran)

Goal: Check whether, for one AI initiative of your own, outcomes and costs can be set side by side on the same basis

Prerequisites

  • One initiative that is reporting outcomes
  • The person in charge of that initiative and someone in finance

Steps

  1. Write down the source and period of the outcome figures
  2. Write the costs of introduction, checking and fixing, operation and data preparation item by item, as far as known
  3. Write "unknown" for items you do not know

What success looks like

  • Outcomes and costs line up over the same period and unit
  • Unknown items are made explicit

When to stop

  • If half or more of the cost items remain unknown, do not expand on the basis of outcome figures alone

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
Mexico (KPMG Mexico)
Region concerned:
Large companies in Mexico and 20 countries worldwide
Source language and what was read:
Spanish / Read the Spanish press release text (the full survey report is unconfirmed)
HAKOMIRO’s check — confirmed conditions for Japan:
For Japanese readers, only an outline of the tendency that there is a gap between self-assessed outcomes and mechanisms to compare with investment could be confirmed.
HAKOMIRO’s check — not confirmed for Japan:
Figures for Japanese companies, the number of respondents in Mexico and the denominators for each question, and the full report are unconfirmed.
Technical details and handoff

The survey covers more than 2,100 executives in 20 countries, and the Mexican portion is only the figures in the press release.

Objective:
Build an evaluation table that sets outcomes and investment on the same basis
Inputs:
Outcome indicators and period; Amount for each cost item
Expected output:
A table setting outcomes and costs side by side over the same period and unit
Constraints:
The Mexican denominator is unknown; Answers are self-reports, not third-party measurements
How to verify:
Check that the cited shares match the press release; Check that global values and Mexican values are not mixed up

New | Work and society

Estée Lauder partners with Microsoft and expands Copilot to more than 14,000 office employees worldwide (no outcome figures)

What it means for work (editorial judgment)
Announcements of AI adoption at large companies tend to mix "expanded use" with "produced results". This announcement covers only the former and shows no usage, productivity or costs. It is an example of separating the number given access from the results of actual use when you read AI adoption reports in your own organization.
What changed
On October 30, 2026, The Estée Lauder Companies announced a strategic partnership with Microsoft. It is investing in an integrated consumer data foundation and extending Copilot to its worldwide office employees (more than 14,000). Employees are described as being at the stage of trying productivity agents, and improved productivity, expanded margins and growth are given as future goals. Microsoft's article reports the same announcement.
Who it concerns
AI adoption leads at large companies, and executives and procurement staff who read adoption reports to make decisions.
What to check or decide today
When reading adoption reports, check "the number given access" separately from "the amount actually used and the results". Read an announcement without outcome figures as a statement of goals, and do not use it as evidence of cost-effectiveness.
Limits
In the company's and Microsoft's announcements, there are no figures on usage, productivity, cost-effectiveness or contract terms. The number who can use Copilot and the number who actually use it are different. Goals are future hopes, not achievements. We have not checked rollout in Japan.
Result
Unknown: The announcement gives only the number who can use Copilot (more than 14,000), and no figures on usage volume, productivity or cost-effectiveness. Granting access cannot be read as a result.
Cost (partial only, not a total)
Unknown: Costs, contract terms and the work required for adoption are not stated.
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Not confirmed in the source
  • Data preparation: Not confirmed in the source
Stage of the effort: Adoption
It says Copilot access was extended to office employees worldwide.
  • What was expected: Improved productivity, expanded margins and growth (the company's goals)
  • What was observed: Not stated
  • Why unknown: The announcement has no figures on usage volume or results, so no observed results can be confirmed.
  • What to check next: Will usage and results be published a year later? How do they compare with the goals?

A small trial to try (the editors’ proposal, not a result we ran)

Goal: Check whether, in your own AI adoption reports, you can read "number who can use it" separately from "results of use"

Prerequisites

  • One adoption report or adoption plan document

Steps

  1. Classify the figures in the document as "number who can use", "actual use", "results" and "cost"
  2. Write out the categories that are left blank
  3. Ask the person in charge what records are needed to fill the blanks

What success looks like

  • A list shows where information exists in the four categories

When to stop

  • If both the results and cost categories stay blank, do not use that document as grounds for expansion

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
United States (The Estée Lauder Companies, Microsoft)
Region concerned:
Office employees worldwide
Source language and what was read:
English / Read the text of the English corporate announcement and the Microsoft article (no translation used)
HAKOMIRO’s check — confirmed conditions for Japan:
For Japanese readers, only an outline of what the adoption announcement says could be confirmed.
HAKOMIRO’s check — not confirmed for Japan:
Rollout in Japan, costs, contract terms and usage are unconfirmed.
Technical details and handoff

Investment in an integrated consumer data foundation and the expansion of Copilot are included in the same announcement. The technical configuration, permission design and data handling are not stated.

Objective:
Organize the facts that can be confirmed from the adoption announcement, and the unknowns
Inputs:
Company announcement; Microsoft article
Expected output:
Table of confirmed and unknown categories
Constraints:
Granting access is separate from actual usage; The technical configuration cannot be known from the primary source
How to verify:
Confirm that the two primary sources are the same announcement; Record as facts only the items that have figures

New | Building and development

AI coding input tokens fall from 92,385 to 5,445 with structured test output (experiment in the author's isolated environment)

What it means for work (editorial judgment)
This is one experiment report suggesting that, when you hand development work to AI, cost and time may depend not only on the AI's intelligence but on the volume of logs you have it read. It is one reason to include formatting of tool output among the items when estimating a development team's running costs.
What changed
On an isolated virtual machine, using the same model and the same task of fixing a payment-processing bug, the author compared passing the raw test output as-is with passing structured JSON containing only the names, lines and diffs of failing tests. Cumulative input to completion (model calls: three for raw output, two for structured) went from 92,385 to 5,445 tokens, estimated cost from about $0.077 to about $0.010, and time from about 25.9 to about 16.9 seconds; the independent tests reportedly passed 20 of 20 under both methods.
Who it concerns
Engineers who use AI coding agents in development, and development leads who decide their running costs and how to proceed.
What to check or decide today
In your own team's agent operation, measure how much test and build output accumulates in the conversation. If you try a mechanism to format the output, check on your own team's tasks that the correct-answer rate does not fall.
Limits
This is an experiment by the author on one task and one model (gemini-3.8-flash in the article). We have not confirmed independent reproduction. It uses an artificially built payment task, so the effect on real business code is unknown. Tokens and costs are estimates in the article.
Result
Cumulative input to completion 92,385→5,445 tokens (model calls 3→2), time about 25.9→about 16.9 seconds, tests 20/20 under both methods (unit: Tokens, seconds | period: The author's single experiment (no period stated) | basis: Measured by the party itself (no third-party verification))
Cost (partial only, not a total)
Estimated cost about $0.077→about $0.010 (unit: US dollars (estimated) | period: One run of the same experiment | basis: Estimate)
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Included
  • Data preparation: Not confirmed in the source
Stage of the effort: Trial / proof of concept
This is a report of the author trying an artificial payment task once on an isolated VM.
  • What was expected: Trimming the output reduces input volume and time
  • What was observed: Input volume down about 94%, tests the same at 20/20 (the author's single measurement)
  • What to check next: Does it fall by the same amount in real maintenance work and with other models? Does trimmed output cause things to be missed?

A small trial to try (the editors’ proposal, not a result we ran)

Goal: In your own environment, measure how much the amount passed to AI falls when test output is trimmed

Prerequisites

  • A development environment where tests can run
  • Records that show the input volume passed to AI

Steps

  1. Measure the output volume once without trimming
  2. Set it up to pass only a summary of failures and run the same work
  3. Compare input volume, time and test results

What success looks like

  • Input volume falls and test results do not change
  • The cause of failures can be read from the trimmed output

When to stop

  • Test results change, or the cause of failures cannot be read
  • The effort of building the trimming exceeds the savings

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
The author's location is unconfirmed from the primary source (an individual's technical article in Japanese)
Region concerned:
The author's isolated virtual machine (no regional limit)
Source language and what was read:
Japanese / Read the Japanese original article text and comparison table
HAKOMIRO’s check — confirmed conditions for Japan:
We checked the steps and figures written in Japanese.
HAKOMIRO’s check — not confirmed for Japan:
We have not checked the availability or cost in Japan of the compression tools or models the article tried.
Technical details and handoff

The author ran an artificial payment task on an isolated VM and compared each method until completion with the same model, gemini-3.8-flash. Raw output took three model calls and structured output two, so it is not a comparison fixed at the same three turns.

Objective:
Measure in your own environment the effect of preprocessing that summarizes test output before passing it
Inputs:
Test execution command; Output before it is passed to AI; Record of input tokens
Expected output:
Comparison table of input volume, time and test results before and after trimming
Constraints:
AI does not misjudge the cause because of missing output; The author's result is from a single experiment
How to verify:
Run the same task several times and look at the variation; Check that the cause of failure remains in the trimmed output

New | Building and development

When passing Xcode operations to AI, the subagent tools setting could not narrow the tools, and they had to be removed with disallowedTools (author's environment)

What it means for work (editorial judgment)
This is one measured report that when you hand development tool operations to AI, assuming that "I listed the tools to use, so it can't use the others" may let unintended writes through. It is grounds for a lead introducing AI development tools to check the tools that can actually be used, rather than relying on configuration alone to narrow permissions.
What changed
In the author's environment (Xcode 27.2, Claude Code 2.1.316), of the 54 tools exposed by Xcode's MCP, writing only 4 in a subagent's tools setting still showed all 54, and XcodeWrite, which was not listed, also went through and created a file. Removing the 50 other than the 4 to be used with disallowedTools left only 4 visible, and preview capture also went through, the author reports. The author runs regular builds and tests with xcodebuild and limits MCP to uses such as preview capture.
Who it concerns
Engineers who have AI agents operate a Mac app development environment, and leads who decide its permission design.
What to check or decide today
If you have AI operate tools in your own organization, do not trust settings you "meant to allow"; actually output the list of tools visible to the agent and check it. When a tool update adds tools, take the list again.
Limits
This was checked in the author's own environment with one combination of versions. The article itself says the tool configuration changes with the Xcode version. We have not confirmed reproduction in other environments or future versions, nor any universal guarantee. The article was published 2026-10-31, and the body was updated on 2026-11-01.
Result
Specifying 4 of 54 tools still showed 54 (XcodeWrite worked). Prohibiting the other 50 left 4 (unit: Number of tools | period: Verification in the author's environment (no period stated) | basis: Measured by the party itself (no third-party verification))
Cost (partial only, not a total)
Unknown: The primary source gives times of 86 seconds for capture and 32 and 31 seconds for other checks. Total effort and total cost including introduction, setup, checking and fixing, and operation are unconfirmed.
  • Introduction and initial setup: Not confirmed in the source
  • Effort to check and correct: Not confirmed in the source
  • Operations: Not confirmed in the source
  • Data preparation: Not confirmed in the source
Stage of the effort: Unknown
This is a verification report in the author's personal environment and does not describe the stage of organizational adoption or operation.
  • What was expected: Not stated
  • What was observed: An allow setting did not narrow it, and a prohibition setting did (author's environment)
  • Why unknown: There is no statement of the adoption stage, and no expected effect is stated.
  • What to check next: Is the result the same on other versions and environments? Did the provider confirm or fix the behavior?

A small trial to try (the editors’ proposal, not a result we ran)

Goal: In your own environment, check that the tools shown to AI are narrowed as intended

Prerequisites

  • An environment that uses the tool integration in question
  • A way to see the list of available tools

Steps

  1. Specify the allowed tools and check the list of tools actually visible
  2. Check that the number of visible tools matches your intent
  3. If not, narrow again with a prohibition setting and recheck

What success looks like

  • The list of visible tools matches the intended scope

When to stop

  • If the list differs from the intent and a prohibition setting cannot narrow it, do not widen the scope in which that integration is used

Basis

Source links use original titles or HAKOMIRO descriptions. Japanese originals retain their published wording.

Publisher’s location:
The author's location is unconfirmed from the primary source (an individual's technical article in Japanese)
Region concerned:
The author's development environment on macOS (no regional limit)
Source language and what was read:
Japanese / Read the Japanese original article text
HAKOMIRO’s check — confirmed conditions for Japan:
We checked the steps written in Japanese and the results on the version as of 2026-10-31.
HAKOMIRO’s check — not confirmed for Japan:
We have not checked behavior on other versions or consistency with official Apple or Anthropic documentation.
Technical details and handoff

The author verified this on Xcode 27.2 and Claude Code 2.1.316. Specifying 4 in a subagent's tools still showed 54, and specifying the other 50 with disallowedTools left 4.

Objective:
Check the setting that narrows permissions with a prohibition list when an allow list does not work
Inputs:
The version in use; Output of the tool list
Expected output:
A setting where only the intended scope is visible, and a record of the check
Constraints:
Behavior may change when the version changes; A report from the author's single environment
How to verify:
Record the version and recheck the list; After setting the prohibition, confirm that unneeded tools cannot be called

What was checked

  • Done: Body text of the UK government (MHCLG Digital) blog, 2026-10-30. On 2026-11-01 we read the full text at the official URL and checked the figures and limitations.
  • Done: Body text of the Cotec (Spain) report page and announcement, 2026-10-26. On 2026-11-01 we read both pages and checked the data period (full-year 2025 and Q1 2026) and the scope (companies with 10+ employees, eight kinds of AI).
  • Done: Body text of the KPMG Mexico press release, 2026-10-28. On 2026-11-01 we read the text. It does not give the number of respondents in Mexico or the denominator for each question, and we have not checked the full report.
  • Done: Body text of the Estée Lauder announcement and the Microsoft article, 2026-10-30. On 2026-11-01 we read both and confirmed that they are the same announcement and that it has no outcome figures. The Microsoft page could be read on re-fetching.
  • Done: Body text of the Zenn article by tappers, 2026-10-31. On 2026-11-01 we read the text and the comparison table. It is the author's own experiment report, and independent reproduction is unconfirmed.
  • Done: Body text of the Zenn article by mochiai, 2026-10-31 (body updated 2026-11-01). On 2026-11-01 we read the text. It is a verification in the author's own environment, and reproduction in other environments is unconfirmed.

Background and past editions

As background, you can read reviews of 15 primary sources from August–October 2026 and a list of 26 materials. These are sources checked in the past, not today's developments.

Machine-readable English version (JSON)