Daily AI Zine
Saturday, August 8, 2026
Issue No. 024 · Tokyo · full edition

✦ Today's Big Thing

Agent work is becoming measurable

GitHub’s metrics move is the useful shift today: Claude and Codex style agent activity inside repos can now be counted, governed, and sold as an operating discipline.

5 min read · 7 sections

In Brief
  1. GitHub added agent app activity to the Copilot usage metrics API.
  2. Enterprise MCP allowlists now give Copilot admins a central way to permit or block connector servers.
  3. Cloudflare is moving bot and agent judgment from one-time risk scoring to continuous trust evaluation.
  4. OpenAI’s researcher access push is a reminder to turn ChatGPT into a structured briefing workflow, not just a chat box.

Today's Big Thing

The one thing that matters

Use it

Agent app activity enters Copilot metrics

The practical shift is not a new model. It is accountability for agent work inside GitHub.

GitHub added agent app activity to the Copilot usage metrics API. That matters because teams can already run partner agents such as Claude and Codex inside GitHub workflows, but agent work is hard to manage if it disappears into individual sessions. The new metric line gives operators a way to ask which repos are using agents, how often they are being invoked, and whether agent spend is linked to reviewable output. For a founder shipping with coding agents, this turns agent usage from a vibe into an operating signal. Start treating each agent run as repo-level production activity, not personal experimentation.

T1GitHub Changelog08/07

My AI Ecosystem

Your actual stack

Use it

MCP allowlists give Copilot admins a hard gate

GitHub now lets enterprise owners centrally control which MCP servers GitHub Copilot clients may run, using allowed and denied server keys in managed settings. MCP is the connector layer that lets AI tools reach external systems. The useful part is governance: teams can stop treating connector access as a user-by-user preference and make it part of workspace policy.

T1GitHub Changelog08/07
Test it

Kimi K3 returns to GitHub Copilot

GitHub says it has resumed rolling out Kimi K3 in GitHub Copilot, billed at provider list pricing under usage-based billing. This is not a reason to switch coding lanes wholesale. It is a reason to keep a small model comparison file for routine tasks, especially bug triage, refactors, and explanation work where cost and latency can matter more than peak model quality.

T1GitHub Changelog08/06
Watch it

Hot signal: Cloudflare starts grading agent behavior over time

Cloudflare says it is shifting bot mitigation from point-in-time risk assessment to continuous trust evaluation for bots and agents. That is a meaningful signal for anyone building agent-accessible websites or workflows. The web is moving toward judging agents by behavior over a session, not only by whether a request looks suspicious at entry.

T1Cloudflare Blog08/07

CoWork Corner

Claude CoWork, day to day

CoWork: make connector approval visible

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change showed up in today’s candidates, so use the GitHub MCP allowlist move as the pattern to revisit. Create one CoWork page called Connector approvals, then record each MCP server, what it can access, who benefits, and whether it is allowed, denied, or pending. The point is not bureaucracy. It is making connector risk visible before another tool is added.

T1GitHub Changelog08/07

GPT Desk

OpenAI, ChatGPT, Codex

Use it

Codex activity now needs a repo ledger

What changed: GitHub’s Copilot usage metrics API now includes agent app activity, including workflows where partner agents such as Codex run inside GitHub. Why Adrian should care: Grey Group OS needs to know which Codex work actually moves repos forward, not just which sessions felt productive. What to do: pick one active repo today, add a simple agent-run ledger with task, model, files touched, review result, and whether the output shipped.

T1GitHub Changelog08/07
Test it

Turn ChatGPT into a research briefing room

What changed: OpenAI is giving 100,000 academic researchers free access to its most advanced ChatGPT models for research, collaboration, and discovery. Why Adrian should care: the useful pattern is not academic access itself, but repeatable research workflows for Goodsense, SET, and Japan-English proposal work. What to do: create one ChatGPT project template that turns raw links, notes, and client context into a two-page briefing with assumptions, unknowns, and next questions.

T1OpenAI News07/29

Small Money Systems

Small, repeatable, real

Test it

Small system: agent usage audit for small teams

System: build a one-page agent usage audit that shows where GitHub Copilot, Codex, Claude, and MCP connectors are being used, what is governed, and what is invisible. Customer: small production, design, or dev teams using AI coding tools without a clear operating policy. Offer: they receive a repo-by-repo usage snapshot, connector risk table, and three cleanup actions. Price: ¥30,000 for the first audit. Existing assets: Grey Group OS, GitHub repos, Claude Code habits, and Adrian’s production workflow judgment. AI workflow: Claude Code or Codex drafts the audit from exported metrics, repo notes, and MCP allowlist rules. First action: make the audit template this hour. Repeatability: reuse the same template monthly. Effort: one hour. Expected value: paid audit revenue and cleaner client delivery.

T1GitHub Changelog08/07T1GitHub Changelog08/07
Test it

Small system: research brief pack for Japan work

System: create a paid briefing pack that turns messy market notes into a concise Japan-English client brief. Customer: foreign producers, founders, or agencies preparing meetings, shoots, or partnership conversations in Japan. Offer: a two-page brief with company context, local assumptions, questions to ask, and a bilingual glossary. Price: ¥20,000 per brief. Existing assets: Goodsense positioning, SET planning habits, Adrian’s Japan business context, and existing proposal templates. AI workflow: ChatGPT structures the research, then Claude Code or Codex can turn the template into a repeatable form and delivery page. First action: draft the intake form today. Repeatability: each brief reuses the same prompt chain and layout. Effort: half day. Expected value: small-ticket revenue and warmer leads.

T1OpenAI News07/29

Build Next

Deployable now

Use it

Build a repo agent-activity ledger

What to build: a small repo dashboard that records each agent app run, task, model, repo, pull request, and shipped result. Why now: GitHub’s Copilot usage metrics API now includes agent app activity, so agent use can be measured instead of guessed. Effort: one hour. Expected impact: less wasted agent spend and faster review of AI-generated work. Dependencies: GitHub access, at least one repo using Copilot or agent apps, and a simple place to store the ledger.

T1GitHub Changelog08/07
Use it

Build an MCP allowlist check sheet

What to build: a simple MCP server inventory with columns for server name, data touched, allowed status, owner, and review date. Why now: GitHub now supports enterprise-managed MCP allowlists for Copilot clients, which makes connector policy concrete. Effort: 15 minutes. Expected impact: fewer risky connectors and clearer handoff when tools change. Dependencies: a known list of MCP servers currently used or planned, plus one shared workspace where the sheet lives.

T1GitHub Changelog08/07

Try This Today

One action, right now

Pick one GitHub repo and write down every agent-assisted task from the past week: tool used, output reviewed, and whether it shipped. If the answer is unclear, that is today’s fix. Make the ledger before adding another agent run.