On 4 November, a Hacker News user asked a simple question: “What's your current AI agentic setup?” Within two days it had more than 500 points and over 600 comments, which makes it one of the larger informal surveys of how developers actually work with coding agents right now.
It is not a scientific sample. People who reply to a thread like this tend to be the ones who enjoy tuning their tools. But that is also what makes it useful for a team lead: it shows where the enthusiasts have landed after a year of experiments, and which habits are worth bringing to everyone else.
The numbers
The AI News Report had three readers go through every comment and tag the tools each person named. Of the 424 commenters who named at least one AI tool:
- 180 named Claude Code, and 103 named Codex;
- 120 named both a Claude product and an OpenAI product;
- 148 named an open-weight or local model, and 58 of those used no Claude or OpenAI tool at all;
- 105 described running agents in parallel, and 56 described a review loop.
Two caveats from the counters themselves: anyone who wrote a bare “Claude” in a coding context was counted as a Claude Code user, which pushes that number up (a plain keyword search finds 141), and the thread kept growing after they counted it.
What stands out
Mixing tools is normal. More than a quarter of the people who named a tool use both Anthropic and OpenAI products, and a third mention open or local models. On a team, that means you should expect several agents in the same repository, not one. Rules, conventions and project context that only one tool can read will be missed by the others.
Parallel work is mainstream among power users. About one in four described running several agents at once, often in separate Git worktrees. One commenter, parkersweb, built a custom skill just for “creating worktrees with the right naming convention”.
Isolation is still improvised. People sandbox agents in VMs, containers, or in one case by creating separate user accounts: FrameworkFred “sandboxed by simply adding additional users to my machine”. It works, but every developer is solving it differently.
Review is becoming a loop, not a step. Several people have one agent review another's work before a human looks. As smashed put it: “I just ask Claude code to do a code review of PR X and it spawns a sub-agent.” Others write a design document first and have the agent critique it before any code exists. chasd00 has Claude write the design, then “carefully review” it before producing an implementation plan.
Simpler often wins. Not everyone is adding layers. thefourthchime went the other way:
“I got rid of all my skills and just set it to medium.” thefourthchime on Hacker News
What a team lead can take from it
A thread full of individual setups is a reminder that most teams don't have a team setup at all. Each engineer has tuned their own, and the results don't add up. A few shared defaults would capture most of what the enthusiasts have learned:
- Write rules once, for every agent. If half your team uses Claude Code and half uses Codex, keep one source of truth and generate the files each tool reads. We wrote about how in CLAUDE.md, AGENTS.md, Cursor rules.
- Agree on how parallel work is done. A worktree naming convention, a cap on open agent pull requests, and a rule for who merges keep parallel agents from turning into parallel conflicts.
- Pick one isolation pattern. Whether it is containers, a dev VM or a separate user, choose one, document it, and make it the default for anything that runs unattended.
- Make the review loop explicit. An agent reviewing an agent is useful, but decide which checks it runs and where a human must still sign off. Otherwise “reviewed” means something different on every pull request.
- Start from a plan. The design-doc-first habit costs minutes and saves hours. Make it the norm for anything bigger than a small fix.
- Resist over-tuning. A setup only one person understands is a liability when they are on holiday. Prefer a plain shared setup that everyone can run.
The takeaway
The individual answers differ a lot, but the direction is clear: several agents, several tools, work in parallel, and review as a loop. The teams that get the most from that are the ones that turn personal tricks into shared defaults, written down where every agent and every engineer can find them.
Sources: Ask HN: What's your current AI agentic setup? (4 November 2026), The AI News Report's count of the thread (6 November 2026).