Community Polls

When a required agent check fails to run, what happens?

A timeout or unavailable service can leave an action unchecked. Think of the security or verification check your agent actually uses. What does the workflow do when that check cannot finish?

Your agent's pull requests: what actually reviews them?

Cursor reported its reviewer bot took average review time from 4.8 minutes to 3.8, and comment acceptance from roughly half to roughly two thirds.

Where does the human check sit in your agent setup?

Anthropic screens every agent action before it runs. Perplexity lets its agents do everything except touch production.

Does your agent sandbox treat GET as read-only?

OpenAI's read-only web access allowed GET and blocked POST. The wiki it landed on accepts page edits over GET, so read-only became read-write with no exploit involved.

How do you find out an agent tool you use has a security fix?

Two Kiro disclosures in two months have shipped without a CVE, so scanners stay quiet.

Your coding agent runs unattended for a week. Is there a spend cap on it?

Cursor, Claude Code and Codex all shipped goals that outlive a session this week. Spend caps did not ship with them.

Claude Code turned auto mode on by default. Are you leaving it on?

Anthropic's study found a classifier caught 89% of dangerous commands against 13.6% for developers reading the prompts.

Do agent config files go through code review on your team?

ChainDrop planted a SessionStart hook in .claude/settings.json and a folderOpen task in .vscode/tasks.json, then committed them across branches.

Should OpenAI release the rogue agent's full activity logs?

Clem Delangue asked OpenAI publicly for the complete logs of the agent that breached Hugging Face.

Do you know what reads your agent's workspace after a run?

Seven bypasses this week all worked the same way: the agent wrote a file it was allowed to write, and something outside the sandbox trusted it later.

After this week's AsyncAPI attack, are you changing how your agents install dependencies?

Attackers compromised AsyncAPI's release pipelines and pushed five trojanized npm packages. Three coding agents shipped guardrails the same week.

Would you route production traffic through a Chinese open model to cut costs?

US firms now send up to 46% of their OpenRouter tokens to Chinese models, at 60-90% lower cost than leading US options.

Would you swap a frontier API for a cheaper or self-hosted model to cut costs?

A cheaper Sonnet tier, a two-agent harness, and an in-house open-weight model all attacked the same bill this week. Where do you land?

Have you turned on your coding agent's sandbox settings?

Three coding tools shipped agent sandboxes this week, all off by default.

After this week, will you keep a fallback model wired up?

A government order took Anthropic's two best models offline for a week with no warning, and the teams that already had an open-weight backup shipped straight through it. Where do you land?

When you pick a coding tool, how much does being locked to one model maker bother you?

Vendors and startups put real money behind the answer this week. Niteshift raised $7M to route between models, and Apple and New API are building the swap into the plumbing. Where do you land?

Would you give an agent its own identity and standing access to your stack?

Microsoft Scout ships with its own governed login and acts on your behalf without asking each time. Convenient, and also exactly how an agent ends up holding access nobody explicitly granted. Where do you land?

After this week, do you still trust AI coding benchmarks?

An independent benchmark from DeepSWE caught a frontier model recovering answers from git history, and flagged double-digit error rates on a popular leaderboard. So where does that leave the scores you have been using to pick a model?

Would you let a coding agent control your machine while it is locked?

OpenAI's Codex can now drive macOS apps on a locked, sleeping Mac, unlocking just long enough to do the work. Convenience, or a security incident waiting to happen?

Would you take a Forward Deployed Engineer role?

Three frontier labs spent the last two weeks turning FDE hiring into their primary enterprise sales motion. The pay is reportedly excellent. The work is embedding at a customer for six months and making their AI investment actually land. Would you sign up?

How is your team handling the "one person teams" pitch?

Coinbase's Tuesday memo cut 14% and rewrote the org chart around engineer-designer-PM hybrids reporting through a five-layer cap with no pure managers. Cloudflare ran the same play two days later and lost 24% of its market cap. Where is your company landing?

GitHub Copilot moves to AI Credits on June 1. What's your move?

Copilot is replacing premium request units with usage-based credits. Legacy plan multipliers jump (Opus 4.7 from 7.5x to 27x) until you migrate. Same monthly bill, very different effective cost. Where's your team landing?

Which new frontier model are you trying first?

Five labs shipped this week. Opus 4.7 on Bedrock, GPT-5.5, DeepSeek V4, Qwen3.6-27B, and ChatGPT Images 2.0. Which one actually made it into your stack?

Which AI desktop surface are you actually using?

All four frontier labs now have first-class desktop presence. Anthropic redesigned Claude Code into a multi-panel app with multi-session support. Google shipped Gemini as a native Mac app. OpenAI's Codex desktop picked up computer use and memory on Mac. Cursor 3 landed a unified agent interface across repos. Which one do you actually spend your day in?

How are you routing between AI models today?

Anthropic's new Advisor Tool lets Sonnet consult Opus only when it hits a hard decision, cutting costs while improving results. It's a formalized version of what some teams have been building ad hoc. We're curious how developers are actually handling model selection in production right now.

Which hidden Claude Code feature from the leak are you most curious about?

Developers catalogued 44 feature flags in the leaked source. Some are practical (parallel agent coordination, persistent memory). Some are less practical (a Tamagotchi with a "CHAOS" stat). Which one caught your eye?

How will the AI lab race shake out?

OpenAI killed Sora to focus on AGI and robotics. Anthropic is going all-in on coding and agents. Google keeps shipping across video, music, quantization, and models because it has the balance sheet to play every lane at once. The question is whether focus beats breadth, or whether Google's ability to outlast and outspend everyone makes the slow-and-steady approach the winning one.

Should OpenAI own Python infrastructure?

OpenAI acquired Astral, the company behind uv (126M monthly downloads) and Ruff. Both stay open source. How do you feel about it?

Should AI review AI-generated code?

Which AI security story concerns you most?

AI security made headlines from every angle this week. A prompt injection hidden in a GitHub issue title compromised 4,000 developer machines through Cline's AI triage bot. Anthropic found 22 Firefox zero-days in two weeks using Claude. A dormant JavaScript worm on Wikipedia hid undetected for two years before a security audit accidentally triggered it. And Vercel's Dubai region failure cascaded globally because middleware deploys to every region. Which one concerns you the most?

Should AI companies refuse government military access requests?

Anthropic told the Pentagon no. OpenAI negotiated a deal with safety conditions. Same week, same government, opposite approaches. Where do you stand?

Which new foundation model are you most excited to experiment with?

Six foundation models shipped in one week. That's not a typo. Which one are you trying first?

Would you push AI-generated code to production from your phone without ever looking at it?

Spotify ships 650+ AI-generated PRs per month. Engineers approve them from their phones without reading the code because they built the Honk infrastructure over 10 years. In the right system, would you do this?

If your AI chatbot started showing ads, what would you do?

Anthropic just spent millions on Super Bowl ads arguing that AI conversations are too personal for advertising. OpenAI says ads keep AI free for billions who can't afford subscriptions. Both have a point. Where do you stand?

How much of your coding workflow is agent-driven now?

Karpathy says he went from 80% manual to 80% agent-driven in weeks. A viral HN post says agents produce "architectural slop." Where do you land?

How do you feel about the impact of AI on junior developer job opportunities?

Stanford found junior dev employment dropped 20% between 2022 and 2025. Meanwhile, AI tool usage keeps climbing. Where do you stand?

How do you feel about AI-generated ads?

How are you adapting to AI's impact on dev work?

The tools are shipping fast. The job market is shifting. Everyone's figuring out their angle. Where are you at?

What do you trust most when using AI coding tools?

Stack Overflow says only 29% of developers trust AI accuracy. Where does your confidence actually come from?

How do you feel about AI IDEs building their own models?

Which new coding LLM are you trying first?

Three flagship models dropped in GitHub Copilot this week alone—GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash. All available now. Which one are you reaching for?

Who's winning the agent infrastructure race?

The Agentic AI Foundation just launched with MCP, AGENTS.md, and Goose under one roof. But the real question is who ends up controlling the standards. Who do you think will lead the agent infrastructure space?

How much of your code is currently generated by AI?

With coding LLMs and IDE agents getting better every week, we're curious how much of the code you write actually comes from AI.

Which AI coding assistant are you using right now?

After this week's showdown between Cursor and GitHub, we want to know: what are you actually using? Takes 2 seconds, helps us write better content for you.