Ask HN: What's your current AI agentic setup?
Developers are rapidly converging on a few dominant tools—especially Claude Code, Codex, DeepSeek, and Qwen—while experimenting with everything from simple chat-based helpers to elaborate “software factories” of coordinated agents. Many describe juggling multiple subscriptions, harnesses (Pi, OpenCode, T3 Code, herdr, Orca, etc.), and custom orchestrators to parallelize coding, research, and personal automation, but also report cognitive overload, QA headaches, and rising token bills. A notable minority either stick to a single lightweight model or avoid LLMs entirely, citing privacy, safety, and skill-atrophy concerns, and there is growing emphasis on sandboxing, local inference, and keeping humans firmly in the review loop.
Subscriptions & models
- Very common: paid Claude (Opus 5.5/Sonnet/Fable) and ChatGPT/Codex tiers; some people juggle multiple $100–$200 plans across providers.
- Others keep it minimal: single $20 tier, or employer‑paid Copilot/Claude/Gemini only; a visible minority pay nothing and stick to free tiers or local models.
- Open-weight models are heavily used: DeepSeek 4.1 Flash, GLM‑5.3, Qwen 3.6/3.8, Kimi, Mistral, Gemma, etc. via OpenRouter or smaller EU providers.
- Common pattern: “frontier” model for planning/review, cheaper or local models for bulk implementation, summarization, or batch jobs.
Harnesses & workflows
- Claude Code and Codex desktop/CLI are the dominant coding harnesses; many say they’re “enough” and improving quickly.
- Alternatives: OpenCode, Pi/oh‑my‑pi, DeepSeek Harness, Antigravity, Kiro, Maki, Hermes, Muse Code, Zed/JetBrains/Cursor integrations, and terminal multiplexers (tmux, Herdr, Rex, Orca, T3 Code).
- Several people have built custom harnesses (often open source) tuned to their own UX: kanban dashboards, card/canvas UIs, Emacs integrations, or Slack/Telegram frontends.
Orchestration, subagents & “software factories”
- Many rely on built‑in subagent features in Claude Code, Codex, DeepSeek Harness, Pi, etc. for fan‑out tasks and code review.
- A long tail uses explicit orchestrators or platforms: OpenClaw, Hermes Agent, Grok Bot, Dots, Muse, ctlflow, Paseo, Gas Town, Kiro‑style factories, etc.
- Quite a few have rolled their own “software factories”: agents driven by Jira/Linear/GitHub issues, kanban boards, or sprint cards; agents open PRs, run tests, perform adversarial reviews, and loop until human approval.
- Others are skeptical of swarms, calling them token‑burning, brittle, or unnecessary compared to a single strong agent plus human steering.
Local vs cloud, cost & performance
- Some run everything locally (Mac MLX, RTX/Strix Halo/DGX Spark rigs) with Qwen/Gemma/GLM/DeepSeek, citing privacy, control, and low marginal cost.
- Others prefer cloud subs, arguing subscription buckets beat API pricing and avoid hardware hassle; a few spend $400–$1000/month, but many operate comfortably on $10–$40.
- Thread notes trade‑offs: frontier closed models are still generally stronger and more concise; open weights are rapidly “good enough” for many coding and automation tasks.
Security, isolation & attitudes
- Isolation strategies range from “raw on my laptop” to elaborate sandboxes: bubblewrap, Docker/microVMs, dedicated Unix users, libvirt VMs, Coder workspaces, or custom gVisor/gateway setups with credential proxies and URL allowlists.
- Some explicitly worry about account bans, safety filters, data retention, or prompt injection; others are more relaxed.
- Opinions on AI itself vary widely: from “promoted me from programmer to architect” and near‑addictive productivity, through cautious but enthusiastic adoption, to people who refuse to use LLMs at all over ethics, safety, or fear of cognitive offloading.