AI Daily Digest · 2026-09-13
🔬 New AI Craft
1. Cognition has Devin test its own work with GPT-6 Astra
Cognition wired GPT-6 Astra into Devin so the agent writes tests, runs them, and fixes failing cases before handing work off — a generate, self-test, repair loop. Unlike an in-session Agent Plan flow that stops at "code written," verification stays inside the same agent, a concrete take on AI-generated unit tests via agent self-verification rather than bolt-on test tooling.
https://openai.com/index/cognition-devin-testing-with-astra
2. Real-SWE benchmarks agents on private, real-world enterprise codebases
withSpecific released Real-SWE, which evaluates coding agents on private production repositories instead of synthesized tasks. Where SWE-bench-style public sets are clean, this measures whether an agent can work inside the messy, partially tested repos most engineers actually own.
https://withspecific.com/benchmarks/real-swe
3. What AI is doing to code review
The Pragmatic Engineer digs into how AI reviewers changed real review workflows: AI handles the first pass while humans focus on architecture and semantics, at the cost of review fatigue and blind acceptance of AI suggestions. A useful reality check on top of the AI code review pipeline you already know.
https://newsletter.pragmaticengineer.com/p/what-is-happening-with-code-reviews
🛠️ Tools & Tips
1. tracecrate
A local-first workbench for agent traces — reads Claude Code, Codex, and OTLP logs, diffs runs, and needs no backend or API keys.
https://github.com/FankChen/tracecrate
2. worktrunk
A git worktree CLI built for parallel AI agent workflows, spinning up isolated worktrees so agents stop fighting over one branch.
https://github.com/max-sixty/worktrunk
3. ai-data-extractor
Exports chat history from Claude Code, Cursor, Windsurf, Aider, and Cline into searchable data for retrospectives.
https://github.com/kruzovic7/ai-data-extractor
⭐ Open Source Highlights
1. YuE2
Frontier music generation with symbolic planning, zero-shot covers, and agentic music editing.
https://github.com/multimodal-art-projection/YuE
2. system_prompts_leaks
A collection of extracted system prompts from Claude, ChatGPT, Codex, and Gemini — handy reference for prompt engineering.
https://github.com/asgeirtj/system_prompts_leaks
3. agent-work-runtime
Persistent work state and minimal context for long-running agents, cutting the need to replay history every turn.
https://github.com/originoneai/agent-work-runtime
📰 Industry News
1. Anthropic CEO Amodei argues it is time to "pace the frontier" and put brakes on AI development
2. Sam Altman says taking OpenAI public in 2026 would be "ill-advised"
3. OpenAI's rogue agent tried to hack another company back in May
4. Robot-training-data startup Mecka AI nears a $500M valuation in a Sequoia-led round
5. Y Combinator's Garry Tan wants US open-weight labs to distill frontier models too
6. Kimi maker Moonshot AI targets $2B in annual revenue
7. Trump administration eases data center pollution rules as AI power and health concerns grow
🚀 Major Releases
1. Salesforce launches Slackforce Surfaces, generating polls, dashboards, and decks inside Slack with AI
2. OpenAI introduces a ChatGPT built for financial services
3. OpenAI expands AI access and cyber defense for federal, state, local, and tribal governments
