Last week we tracked MAI-Code-1-Flash becoming the default GitHub Copilot model and Intune’s MCP kill switch. This week is about consolidation and multiplication simultaneously. Microsoft is merging four Copilot experiences into one super app. Copilot Studio shipped its most significant overhaul ever. M365 Copilot’s model picker now spans four frontier models across two providers. And GitHub Copilot CLI has matured into a genuine agent platform.
For IT leaders, Copilot stops being a collection of products and starts being a single platform with multiple entry points.
1. Copilot Super App: Four Experiences Become One
Satya Nadella confirmed on Microsoft’s FY26 Q4 earnings call (July 29) that Copilot Chat, GitHub Copilot, Copilot Cowork, and autonomous AutoPilot agents are merging into one unified application — target release by end of September 2026. Users will toggle between personal and work contexts within the same interface.
A new paid AutoPilot tier introduces background autonomous agents that work without constant prompting. The first, Microsoft Scout, handles scheduling, meeting prep, and deadline tracking across Teams, Outlook, OneDrive, and SharePoint. Pricing is not yet disclosed.
Copilot Podcasts and consumer Deep Research are retired August 18 — podcasts will be deleted with no export option. Copilot Labs is also discontinued. M365 Premium subscribers redirect to the Researcher agent for in-depth reports.
What this means for IT leaders: The super app launch is weeks away. Save any Copilot Podcasts content before August 18 deletion. Migrate consumer Deep Research workflows to the enterprise Researcher agent.
2. Copilot Studio’s Agent-First Overhaul and the GitHub Copilot Harness
Copilot Studio shipped its most significant redesign on August 3, moving from chatbot-first to agent-first. The old architecture (question → topic → predefined flow → response) is replaced by an orchestrator-based model (question → orchestrator → knowledge, tools, skills, workflows → response). Topics are gone; instructions are paramount.
Three harnesses now power agents. The Standard harness handles rule-based agents, covered by M365 Copilot licenses for B2E. The Copilot Chat harness extends M365 Copilot with enterprise knowledge. The new GitHub Copilot harness (GA August 3) enables long-running, goal-seeking agents that plan, act, and adapt across steps.
The GitHub Copilot harness billing model is a paradigm shift: billing starts when you build, not when you publish. Natural-language authoring, preview conversations, and evaluation runs all consume credits. Runtime tasks range from 100–300 credits ($1–$3) for light operations to 500+ credits ($5+) for heavy multi-step processes — 5–15× a Standard harness run. Harness choice is permanent; agents cannot transfer between harnesses.
The credit system pools at tenant level across Cowork, Copilot Studio, Work IQ APIs, Dynamics 365, and Power Platform. One credit equals $0.01, with pay-as-you-go or annual pre-purchase plans (5–20% volume discounts). At 125% of prepaid capacity, custom agents are disabled.
What this means for IT leaders: Budget forecasts must include development, testing, and evaluation — not just runtime. Recreate Copilot spending alerts (a Cost Management update wiped existing ones). Choose harnesses carefully; the decision is irreversible.
3. Multi-Model M365 Copilot: Four Frontier Models
The M365 Copilot model picker now includes GPT-5.6 (OpenAI), Claude Sonnet 5, Claude Opus 5, and Claude Fable 5 (Anthropic). GPT-5.6 is the preferred default across Word, Excel, PowerPoint, Chat, and Cowork. Claude Sonnet 5 handles agentic multi-step work. Claude Opus 5 is available in Word and Copilot Chat — strong but token-hungry. Claude Fable 5 targets longer complex projects in Cowork.
Admins enable Anthropic models per user or group in the M365 admin center. Critical: OpenAI-operated models from GPT-5.6 run under OpenAI as a subprocessor, and the admin toggle was auto-enabled July 24 unless already set to “No users” (MC1422074). Tenants with EU Data Boundary concerns should check this immediately.
Key features landed alongside. Power BI grounding lets Copilot reason over governed enterprise data with row-level security respected. Word Track Changes rewrites show every Copilot edit as a tracked change. SharePoint live dashboards generate interactive HTML reports connected to source data. Copilot Notebooks are now available to all commercial and EDU customers with OneNote sync.
What this means for IT leaders: Multi-model is the new normal. Develop model-selection guidance — cost varies significantly across models. Check the OpenAI subprocessor toggle if EU data residency matters. Power BI grounding eliminates export-based workflows.
4. GitHub Copilot CLI: A Real Agent Platform
GitHub Copilot CLI has evolved from a terminal assistant into a full agent platform. It manages concurrent sessions, creates isolated Git worktrees, and restores state with /rewind. The Copilot SDK went GA June 2 in six languages, exposing the same execution loop as a programmable layer. The Microsoft Agent Framework integration shipped stable v1.0 on August 4 for .NET and Python with deny-by-default permissions.
Agent Skills are the extensibility standard — reusable instruction bundles with a SKILL.md entry point working across 30+ platforms. The directory lists ~2,100 published skills. Code Review with Agent Skills and MCP servers went GA July 29.
Enterprise governance matured significantly. Managed settings enforce configuration across CLI, VS Code, the Copilot app, and cloud agents. Agent session streaming (public preview) sends 48 hours of session data to SIEM or Purview. OpenTelemetry export provides centrally managed telemetry.
What this means for IT leaders: GitHub Copilot CLI requires the same governance as any development tool. Configure managed settings, enable session streaming, and establish skills review processes.
5. September 1 Deadlines and Model Deprecations
Multiple hard deadlines hit September 1, 2026:
- Six GitHub Copilot models retire: Gemini 3.1 Pro, Claude Opus 4.5/4.6, Claude Sonnet 4.5/4.6, and Raptor Mini. Pinned configurations and CI workflows referencing these will break.
- Claude Sonnet 5 introductory pricing ends — standard pricing ($3/$15 per million tokens) replaces $2/$10 the same day.
- Pre-August 3 Copilot Studio agents on the GitHub Copilot harness move to credit billing.
- Passkeys become default for Entra ID. SMS/voice MFA retires February 1, 2027.
What this means for IT leaders: Audit pinned model configurations now. Migrate to recommended replacements before September 1. Run the entra-sms-voice-usage-analyzer inventory script and build a passkey migration plan.
Strategic Next Steps
- Prepare for the super app — brief stakeholders on the September unification. Save Copilot Podcasts content before August 18 deletion
- Choose Copilot Studio harnesses deliberately — the GitHub Copilot harness bills from first build and cannot be reversed. Start with Standard for proof-of-concept, then graduate with budget forecasts
- Recreate Copilot spending alerts — the Cost Management update wiped them. Set per-policy alerts and monitor credit consumption
- Check the OpenAI subprocessor toggle — verify the auto-enabled setting if EU Data Boundary matters (MC1422074)
- Audit pinned GitHub Copilot models — six retire September 1. Update configurations and CI workflows now
- Enable Power BI grounding — the integration respects row-level security and eliminates export-based workflows
- Configure GitHub Copilot enterprise governance — deploy managed settings, enable session streaming, establish skills review
- Plan passkey migration — inventory SMS/voice users and use the opt-out API only as a temporary bridge
The story this week is convergence. Four apps becoming one. Multiple models through a single harness. Copilot Studio’s agent-first architecture merging building and running. GitHub Copilot CLI’s SDK converging terminal, IDE, and programmatic development on a shared engine. The mandate is clear: stop treating Copilot as point features and start governing it as a unified AI platform.
Follow along at https://x.com/kkaminsk for daily Copilot and Microsoft AI analysis.