GitHub argues chat is the wrong AI UI: Canvas lets your agent build tools instead of burning tokens
GitHub Blog AI/ML · rss · 2026-09-25
GitHub's official blog argues that while chat is the default interface for LLMs, it's often the wrong one—once you know your task, you need a custom UI built for it. The piece showcases Canvas in the GitHub Copilot app: a full-stack mini-app that runs without browser chrome, communicates bidirectionally with the agent, can call third-party APIs, and execute code locally.
Examples include:
- A Connect 4 game you play against the agent (author beat Luna without reasoning, lost to high-reasoning Sol)
- A Winget UI with zero AI inside—which is the point
- A SQLite canvas with IntelliSense, so you stop asking the agent to query for you
- Reviving Windows Live Writer for Jekyll blogging
The core argument: when chat is the only interface, you default to making the agent do everything, which is pure token waste. Better to have the agent build a tool once so all future interactions are free—stop asking GPT-5.6 Sol Max to "stage and commit".
The post also walks through automating a full dev workflow (Research → Prototype → Plan → Implement → Iterate → Finalize) with a canvas, keeping yourself out of the loop except for reviews. The SQLite canvas is one-shottable; the workflow canvas took most of a day to get right.
More from coding & agent
- Why inference speed will be AI products' biggest competitive edge in 1–2 years — itsOmSarraf_ · 2026-11-11
- What happened when I moved vector retrieval into Snowflake SQL instead of orchestrating RAG outside — dima806_dima · 2026-11-11
- OpenAI engineer runs Codex for 26 days straight hunting Atlantis, spawning "vibe archaeology" — Polymarket · 2026-11-11
- He replaced a $5,000/month ghostwriter with 11 Claude Skills and earned $47,000 from one post — aryanXmahajan · 2026-11-11
- Refero Styles ships 2,000+ DESIGN.md files to give Claude Code real design taste — bbssppllvv · 2026-11-11
- Agent circuit breaker: openclaw shorts bot halts itself after 3 failed experiments — Rich_Standard_1724 · 2026-11-11