Ask HN: What's your current AI agentic setup?
A community discussion thread asks working developers how they use AI day to day. Its poster lists two questions: which AI subscriptions people personally pay for, such as Claude, ChatGPT or Codex, and Cursor, and whether they use models outside or instead of those plans. The poster says nobody needs to answer every question, so the replies are informal, self-reported and anecdotal rather than a survey.
The replies describe a mix of setups rather than one standard stack. One commenter pays $40 a month for OpenCode Go and $20 for Claude, and spends between $0 and $100 on DeepSeek's API. Another holds a $14 monthly Z.ai coding plan. Others reach open-weight models such as Kimi, GLM and Qwen through OpenRouter or OpenCode, and at least two run inference on local hardware, including NVIDIA DGX Spark machines. Some have built their own agent harnesses, and at least one runs its agent in a container with restricted file access.
Because the thread has no sampling method, its replies show direction rather than prevalence. The commenters appear to be experienced users who tinker with hardware and low-cost plans, so their spending should not be read as typical. For measured adoption, the developer surveys described below are a better guide.
Why it matters
For companies that sell AI products, the thread shows how quickly developers can move between model providers. Price is a lever: one commenter plans to move a security product toward cheaper open models, partly to avoid refusals. Stack Overflow's 2026 survey found that 38% of organizations set usage limits and 27% use different models for different work. Products tied to a single model vendor carry margin and availability risk when buyers can swap in alternatives.
For investors, the value may sit in the harness, routing and cost-control layer rather than only in the base model. Developers are assembling their own orchestration, and employers are monitoring token spend, which favors tools that show usage, enforce limits and route tasks to suitable models. Founders should expect buyers to test open-weight models before committing, and should price products with token costs that customers now watch closely.
Key points
- 1JetBrains surveyed more than 15,000 professional developers between May and July 2026.
- 2JetBrains reported that 90% of professional developers use AI coding agents at work weekly, with 68% using them daily.
- 3Claude Code reached about 39% of professional developers at work in May to July 2026, up from 18% in January.
- 4Stack Overflow's 2026 survey found 38% of organizations set usage limits and 27% assign different models to different work.
- 5
