Copilot, current shape

Before you build on Copilot, know its current shape. Agent mode is generally available^⊙ , and the product has reorganized around agents rather than chat: local agents that run in your editor, a cloud agent that runs in a sandbox and opens a pull request^⊙ , subagents that a lead agent can fan work out to in parallel^⊙ , and a sessions hub to watch them. The October 2025 reorganization brought third-party agents from several vendors into Copilot subscriptions under one roof^⊙ . Custom agents — the definitions you will write next lesson — are honored across the editor and the cloud agent alike.^⊙

You do not need to use all of it. You need to know it exists, because the harness leans on three pieces: local agents for interactive work, subagents for delegation, and the cloud agent for work you can hand off and walk away from.

Choosing a model

Copilot lets you pick a model per request, which means model choice is now a skill. A defensible policy — offered as a policy, not a prescription:

• A capable, fast model as your daily driver for most agent work. Claude Sonnet 5 reached general availability at the end of June 2026^⊙ and is a reasonable default.
• A stronger model for planning- and synthesis-heavy work — writing the data model, reconciling findings, designing the demo script. Claude Opus 4.8 has been GA since late May 2026^⊙ and earns its cost on exactly those tasks.
• A fast, cheap model for mechanical edits — renames, reformatting, boilerplate. Several capable models sit in this tier^⊙ .

The point is not the specific names, which will rotate. The point is to match the model to the job instead of defaulting to whichever felt best last month.

The honest paragraph on cost

Premium requests are metered by a per-model multiplier against a monthly allowance, and the multipliers change^⊙ . Any specific number this manual printed would be wrong by the time you read it, so it prints none — check the pricing page before you quote a figure to anyone. The discipline that survives the churn: reach for the expensive model when judgment is the bottleneck, and the cheap one when typing is. A harness that burns its budget on mechanical edits has its model policy backwards.

In the field

Add three lines to your engagement instructions naming your default model, your planning model, and your mechanical-edit model — by role, not just by name, so the note still makes sense when the names change. You now have a policy instead of a habit.