AI Coding Assistants

Today's Stories

1. 1

   AI to Sort Trash in Sabae City Starting Sept 1

   Sabae City launched a web app on September 1 that uses AI to identify and sort trash when users photograph it with a smartphone. The app can distinguish multiple items in a single photo and provides sorting instructions within about 10 seconds. Last year, fires caused by improperly disposed lithium-ion batteries in non-burnable trash occurred repeatedly at the city's waste facility. The city hopes the AI tool, accessible via its website, will reduce such sorting errors and improve safety.

   The service is free and available through Sabae City's homepage. It is unclear if the city will expand the app's capabilities or monitor its effectiveness in reducing mis-sorting incidents.

2. 2

   Open-source InsForge boosts AI agent backend accuracy

   Agents fail at backend tasks like auth and database setup. InsForge adds a semantic layer with structured, machine-readable capabilities. It delivers roughly 2x more accuracy than Supabase MCP, and 30% better token efficiency. Backend platforms like Firebase and AWS were built for human clicking, so agents get fragmented context and hallucinate. InsForge fixes this at the infrastructure level, letting agents reason about the backend as one system.

   InsForge works with Cursor, Claude Code, Windsurf, and Codex, and is open-source under Apache 2.0. A test built a full ChatGPT clone with Claude Code with no manual configuration.

3. 3

   Meta unveils Muse Spark 1.3, but top results limited

   Meta unveiled Muse Spark 1.3 yesterday, claiming frontier performance and calling it the 'biggest jump' yet in coding and agentic work. The version developers can access now shows significant gains over last month's 1.2 release, especially on long-running agent tasks. The model is one of the strongest price-performance offerings near the top of independent model rankings. However, Meta's best benchmark results come from a 'max reasoning' configuration that is still in safety testing and not broadly available to developers.

   Meta says the max reasoning version will arrive 'shortly.' The third-party benchmarking firm Artificial Analysis evaluated it in a limited partner preview and currently lists no API provider at that configuration.

4. 4

   OpenAI cuts ties with Cursor over Musk's SpaceX deal

   OpenAI announced it would wind down its partnership with Cursor, the AI coding tool startup, because SpaceX—now Cursor's owner after a $60 billion acquisition—couldn't be trusted to comply with its terms of service. OpenAI cited its experience with Elon Musk's companies violating contracts. Cursor was one of OpenAI's top five customers by revenue at the start of 2026, projected to bring in over $1 billion in annualized revenue. Despite this loss, OpenAI is preparing for a public offering next year and reportedly generates over $40 billion in annualized revenue, lessening the financial blow but risking developer trust.

   Cursor's CEO Michael Truell claimed OpenAI models serve only about 5% of Cursor user traffic, a figure OpenAI's Thibault Sottiaux challenged. Anthropic will continue to offer Claude models in Cursor, though Anthropic's dependence on SpaceX for $45 billion of data center capacity may influence that decision.

5. 5

   AWS shows AI-driven coding with two reference builds

   AWS published reference implementations for an AI-driven development lifecycle. One turns SQL schemas into Mermaid ER diagrams. The other runs automated code security checks. The systems keep humans in control while AI does routine work. They show how to add AI to coding without removing human review.

   Both use Amazon Bedrock AgentCore and Claude Sonnet 4. You can find full deployment steps in their GitHub repos.

6. 6

   GitHub Copilot app: Parallel AI agents on one project

   GitHub Copilot app lets you run multiple AI agent sessions at the same time on the same project, each in its own Git worktree, so they don't interfere with each other. This means you can start several tasks at once instead of waiting for one to finish, like doing multiple loads of laundry at the laundromat. You spend your time reviewing results, not babysitting the agents.

   The app lets you start two small tasks simultaneously as a low-stakes way to try it. Each session keeps its own context, so you can switch between tasks without re-explaining.