Explore what to build next

The payment webhook that demoed perfectly and exploded three days after launch: four correct postures for having AI write tests, three high-voltage lines requiring human review (money/permissions/deletion, concurrency, migrations), and a minimal pipeline for tonight.

At MongoDB.local NYC on September 29, MongoDB and Cognition launched Devin for MongoDB Modernizations: Devin rewrites business logic and data-access layers while AMP's deterministic tooling migrates and validates data — early tests claim five-to-six-hour jobs compressed into one. The most down-to-earth form of agents landing: selling migration outcomes, not tools.

Nine of ten first vibe-coding projects die of scope creep. The cautionary tale of "the AI Notion+Slack" archived in week three, the "demoable in a week" rule, four knives for cutting features, and five stall-warning signals: the goal isn't big, it's done.

GitHub announced September 24: a new default policy for Copilot Business/Enterprise features takes effect October 22 — all unconfigured GA features will follow the global default, which is Enabled. With MCP servers in scope and two model retirement waves in October, enterprise governance shifts from approval-based to patrol-based.

Losing more money with every user is the most expensive bug in AI products. The arithmetic of per-user gross margin: why you must price to the P90 user, not the average; the photo app burned alive by its own virality; three cost levers and three margin red lines.

On September 22, JetBrains CEO Kirill Skrygan announced JetBrains Air: a system of products for agentic software development spanning individuals, teams, and organizations — in-IDE agent orchestration, reusable Air Teams workflows, and Air Governance. After 26 years on the individual developer workbench, JetBrains is betting on verification, not generation.

How many "god-tier prompts" are rotting in your bookmarks? The third time you paste the same prompt, it should graduate from a one-off instruction into an installable Skill. The real difference between prompts and Skills, three upgrade signals, and a hands-on skill directory walkthrough.

Anthropic's September 29 red-team report finds Zhipu's open-weight GLM-5.3 built working exploits in 50 of 410 ExploitBench attempts, nearing the unreleased Claude Mythos Preview (56) — while stripping its safeguards costs only about $4,400, and abliterated weights are already circulating. In the open-weight era, safeguards are becoming a suggestion, not an iron law.

The demo is 10%; shipping is the other 90%. A pre-launch checklist built for AI-generated projects: security, data, observability, performance & cost, release & rollback — because agents have three systematic blind spots, and 'shippable' judgment is yours.

Anthropic sent ~950 AI agents to search a massive DNA database; they wrote their own code, sifted 200,000 candidates, and found a novel CRISPR-like enzyme system. The leap from 'executor' to 'explorer' — and three caveats before you get excited.

After a year of AI coding, the bottleneck moved from 'writing code' to 'saying clearly what you want.' A seven-section PRD template for agents, three real bad-vs-good cases, and four advanced techniques — plus why PRD skill is replacing coding skill as the scarce ability.

At DevDay 2026 on September 29, OpenAI unveiled Dots — always-on agents with their own cloud computers, powered by GPT-6 Astra, connecting to 4,000+ apps. Three weeks after Meta's Muse dethroned ChatGPT in downloads, the agent war has entered the 'form factor' phase.