Skip to content
All issues
Published issue

🍻 Agents Got Rules

The model was not the story this week, who gets access was. GitHub split Copilot controls by team, Salesforce started switching on a coworker for eligible seats, and old taps kept disappearing. The seasonal change is upon us.

01

Quick Pour

Friday in five minutes

🍻 AI grew a control panel.

GitHub shipped team-level settings, comment-triggered automations, and per-task reasoning controls for Copilot. Salesforce began a rolling Agentforce Coworker auto-enable for eligible seats. The useful move is not another demo: write down who can run what, against which data, for how much, and who can stop it.

Covered first-party changes · July 31–August 6
MonTueWedThuFriSatSun
Jul 270000200
Aug 33210000
Jul 31
GitHub opened team model policies and posted September deprecations
Aug 3
GitHub added team settings, comment triggers, and reasoning controls
Aug 4
Salesforce began Coworker rollout and GitHub retired its billing preview
Aug 5
Anthropic retired Claude Opus 4.1 on its operated platforms

Only the first-party product changes covered here—not every AI release this week. Salesforce’s August 4 change is a rolling rollout.

02

This Week’s Flight

Lead pour · Governance by team

Copilot split the house rules by team.

GitHub Enterprise admins can now keep company-wide settings fixed while letting selected Copilot settings vary by team. The controls apply across VS Code, Copilot CLI, the Copilot app, and Copilot cloud agent. Operator lesson: lock the floor centrally, then grant exceptions by job and training—not by whoever asks loudest.

Read GitHub’s team settings release (opens in a new tab)

Now pouring · Comments become triggers

One comment can put an agent on shift.

Copilot cloud agent automations can now start from a chosen issue or pull-request comment. GitHub suggests documentation, error investigation, and follow-up tasks. Treat the trigger phrase like a button: make it specific, keep the first action reversible, and leave a visible trail.

Read GitHub’s automation release (opens in a new tab)

Cost control · Think per task

More thinking now comes with a visible tab.

Paid Copilot users can choose the reasoning level when starting a cloud-agent task on supported models. GitHub says higher reasoning can help complex work but uses more tokens and credits. Put the expensive pour on hard problems, not routine cleanup.

Read GitHub’s reasoning control note (opens in a new tab)
03

By the Numbers

Derived comparison · Public notice runway

Three changes gave operators less than a month.

These calendar-day counts run from the first-party announcement or article date to the effective change date.

Public notice to effective change · Calendar days · Derived
Salesforce Coworker · Jul 9 → Aug 4
26 days
Copilot Billing Preview · Jul 7 → Aug 4
28 days
Gemini in Copilot · Jul 2 → Jul 31
29 days
Claude Opus 4.1 API · Jun 5 → Aug 5
61 days

Three recent changes landed with 26–29 calendar days of public runway; Anthropic’s model retirement had 61.

Derived from first-party notice and effective dates. This compares timing, not impact or vendor quality; publication is not proof every administrator saw the notice. Salesforce’s change is a rolling rollout, and Anthropic-operated versus partner-operated model schedules can differ.

04

Smokin’ Hot Skill

Smokin’ hot this weekCommunity-created · Hosted by GitHub · Not paid

GitHub’s Awesome Copilot catalog / Skill on tap

Agent Governance

Use it to sketch the boring parts that keep an agent useful: tool allowlists, review gates, call limits, and an audit trail. Start with one agent that touches customer data or production systems.

Best for
Turning agent permissions into a reviewable policy
Monday move
Require human review for one high-impact tool

Community-created, hosted in GitHub’s catalog, and selected independently by Studio Tokyo. Not paid. Review the skill before use; this issue does not install or execute it.

Open the skill (opens in a new tab)
05

Try This Monday

One useful move · 30 minutes

Give one agent a closing checklist.

Pick the AI tool closest to customer data, production code, or money. Write one page that names its access, triggers, limits, owner, and stop control.

  • List who can start it, which tools and data it can touch, and where every run is recorded.
  • Set one spend or call limit, one human-review gate, and one named person who can shut it off.
06

Last Call

Deadline · September 1

Six Copilot models have a September last call.

GitHub says Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Claude Sonnet 4.5 and 4.6, and Raptor Mini leave Copilot September 1. Annual individual subscribers keep Sonnet 4.6. Enterprise admins may need to enable replacement models before moving workflows.

Read GitHub’s model deprecation notice (opens in a new tab)