Daily AI Zine
Wednesday, July 29, 2026
Issue No. 015 · Tokyo · full edition

✦ Today's Big Thing

Copilot becomes a wider agent bench

GitHub added another reasoning model to Copilot and gave JetBrains users more control over MCP-connected agent flows.

5 min read · 7 sections

In Brief
  1. GitHub Copilot added Grok 4.5 as another fast, agentic coding lane.
  2. Copilot for JetBrains now connects MCP servers and custom agents inside Claude agent flows.
  3. GitHub Actions now holds potentially malicious workflows for approval in public repositories.
  4. Hot signal: Google added more Managed Agents capability in Gemini API, including 3.6 Flash and hooks.

Today's Big Thing

The one thing that matters

Test it

Copilot is turning into a multi-model agent workbench

The useful shift is not only another model. It is Copilot becoming a place to route coding work across models, agents, MCP servers, and IDE controls.

GitHub added Grok 4.5 to Copilot, positioning it for fast agentic coding and complex multi-step workflows. Separately, Copilot for JetBrains added improved OpenTelemetry configuration, model management, MCP server connections, and custom agents in Claude agent flows. The practical point: coding work is moving from one assistant to a managed workbench where the operator chooses model, context, tools, and visibility. For Adrian, this is a test lane, not a migration. Pick one contained repo task, run it through the new Copilot lane if available, and compare it against the current Claude Code or Codex path on speed, review burden, and fix quality.

T1GitHub Changelog07/28T1GitHub Changelog07/28

My AI Ecosystem

Your actual stack

Use it

GitHub Actions adds a safety pause for risky workflows

GitHub Actions now holds potentially malicious workflows for approval in public repositories. The change targets attacks where compromised credentials push workflow files that try to steal CI/CD credentials. Treat this as a reminder to audit which repos can run Actions automatically, especially around deploy keys, Netlify deploys, and token-bearing jobs.

T1GitHub Changelog07/28
Watch it

Copilot for JetBrains gets MCP and custom-agent controls

GitHub says Copilot for JetBrains can now connect MCP servers and custom agents in Claude agent flows, while giving more control over telemetry and model management. MCP means a standard way for AI tools to reach external systems. This matters if agent work needs logs, repos, issue data, or private tools inside a controlled IDE workflow.

T1GitHub Changelog07/28
Watch it

Hot signal: Gemini Managed Agents adds more production controls

Google announced new Gemini API Managed Agents capabilities, including 3.6 Flash and hooks. Hooks are checkpoints where a developer can make the agent run extra logic, such as logging, approval, or cleanup. This is not Adrian’s primary stack, but it is a useful signal: hosted agent platforms are competing on control surfaces, not only model quality.

T1Google AI Blog07/28

CoWork Corner

Claude CoWork, day to day

CoWork should act as the project control room

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 Tier 1 candidates, so treat this as a workflow reset rather than news. Revisit one pattern: keep a single project record with the goal, current assets, open decisions, and the next delegated coding task. Then send only that task to Claude Code or Codex, not the whole messy thread. Anthropic’s agentic coding guidance keeps pointing back to scoped tasks and clear context; CoWork should be the place where that context is maintained, not where every agent improvises.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Use OpenAI’s scientist-agent pattern for internal tooling

OpenAI published a field report on scientists using AI coding agents to modernize scientific computing and accelerate software development in genomics and beyond. Adrian should care because the same pattern fits Grey Group OS: old spreadsheets, proposal fragments, production trackers, and research notes can become small maintained tools instead of one-off files. Do one concrete experiment today: ask Codex to turn one recurring manual calculation or checklist into a tiny reviewed utility, then log whether it saves time or just creates review work.

T1OpenAI News07/28

Small Money Systems

Small, repeatable, real

Use it

Sell a GitHub Actions safety mini-audit

System: build a GitHub Actions workflow-risk mini-audit. Customer: small companies, agencies, and content businesses that have public repos, GitHub Actions, and deploy tokens but no security owner. Offer: a one-page risk report, one cleaned workflow, and a simple approval rule. Price: ¥35,000. Existing assets: Grey Group OS, GitHub, Netlify habits, Claude Code, and Codex. AI workflow: Claude Code or Codex reads workflow files, flags token exposure and risky triggers, then drafts fixes for human review. First action: pick one repo and create the audit template today. Repeatability: the same checklist sells across every client repo. Effort: one hour. Expected value: paid setup work plus reduced chance of a broken or compromised deployment.

T1GitHub Changelog07/28
Test it

Package a coding-agent routing card

System: create a paid coding-agent routing card for teams confused by Copilot, Claude Code, and Codex choices. Customer: founders or small product teams shipping web tools, internal dashboards, or campaign microsites. Offer: one test issue run through selected lanes, a verdict on which lane handles planning, implementation, and fixes best, and a reusable prompt card. Price: ¥25,000. Existing assets: Grey Group OS, Codex, Claude Code, GitHub, and real shipping patterns from production work. AI workflow: run the same contained task through the available tools, compare outputs, and summarize review burden. First action: use Grok 4.5’s Copilot arrival as the trigger for a fresh benchmark. Repeatability: every new model release refreshes the same product. Effort: half day. Expected value: small consulting revenue and sharper internal routing.

T1GitHub Changelog07/28

Build Next

Deployable now

Use it

Build a workflow approval register for active repos

What to build: a small register that lists every GitHub Actions workflow, its trigger, whether it touches secrets, and who approves changes. Why now: GitHub is now holding potentially malicious workflows for approval, which makes workflow trust a live operations issue. Effort: one hour. Expected impact: fewer risky deploy paths and faster review when a workflow changes. Dependencies: existing GitHub repositories, GitHub Actions access, and a human owner for approval decisions.

T1GitHub Changelog07/28
Test it

Run a Grok 4.5 Copilot comparison task

What to build: a one-page comparison sheet for one repo issue, scoring Grok 4.5 in Copilot against the current coding-agent lane on plan quality, patch quality, and review time. Why now: Grok 4.5 has just started rolling out in GitHub Copilot for agentic coding work. Effort: one hour. Expected impact: better model routing and less time wasted sending tasks to the wrong assistant. Dependencies: Copilot access to Grok 4.5 and one small issue with clear acceptance criteria.

T1GitHub Changelog07/28

Try This Today

One action, right now

Open one active GitHub repo and inspect its Actions workflows. Write down the trigger, whether it can access secrets, and who should approve changes. If the answer is unclear, create the approval register from today’s Build Next item.