Skip to content
AI Coding

Ask HN AI Agentic Setup: Practical Starter Guide

The Ask HN: What's your current AI agentic setup? thread is exploding. Here's a beginner stack distilled from real replies - not the $500/mo chaos.

6 min readBeginner

Two reactions when Ask HN: What’s your current AI agentic setup? takes off: mirror the loudest stack (stacked Max tiers, $500 Codex lines, custom swarms), or pull the patterns that keep repeating and ship something usable this week. Beginners win on path two. Path one spends the budget before you know what “done” even means.

The thread crossed into the high hundreds of points fast (~289 pts / ~288 comments in the early window). Replies swing from “I type a prompt and read the output” to multi-VM sandboxes and Telegram-driven Claude loops. Treat that page as a snapshot, not a shopping list – half the comments are power-user theater.

What the Ask HN AI agentic setup thread actually agrees on

Strip the flex and the list is short. Narrow jobs beat god-mode agents. Plan lives in a file before code lands. Handoffs sit in markdown – plan.md, changes.log, notes.md. A second model or a fresh pass reviews. Draft PRs beat silent merges. Tiers matter: cheap/fast for bulk, stronger models for hard design or security.

Names that show up most: Claude Pro/Max with Claude Code, ChatGPT Plus/Pro for Codex, OpenCode (free client; you pay the model), plus OpenRouter or DeepSeek for overflow. Local models appear for privacy and cost – not as day-one gear.

Hands-on: starter agentic setup in one afternoon

Pick one primary coding agent. On a single paid seat, Claude Code on Pro is the cleanest path – $20/mo ($17 annual) per Anthropic’s pricing as of the current public page; Max 5x is $100/mo and Max 20x/Ultra $200/mo if you later hit walls. Want model choice and open source? OpenCode. Already on ChatGPT Plus? Codex CLI is fine. Three installs on day one is how the thread bias sneaks into your wallet.

1. Install and sign in

# Claude Code (macOS / Linux / WSL)
curl -fsSL https://claude.ai/install.sh | bash
cd your-project
claude
# same account as your Pro/Max plan

OpenCode alternative (install docs as checked late 2025/early window on their v2 script):

curl -fsSL https://opencode.ai/v2/install | bash
opencode

claude --version or opencode --version. Command missing? New shell so PATH catches up – mine needed that once.

2. Write AGENTS.md (shared brain)

Drop AGENTS.md at the repo root. Short: commands the agent can’t guess, conventions, hard boundaries.

# AGENTS.md
## Commands
- Install: npm install
- Dev: npm run dev
- Test: npm test
- Lint: npm run lint

## Rules
- TypeScript strict; no any
- Don't edit .env or secrets
- Open a PR; never force-push main
- After edits: run tests before claiming done

## Definition of done
- Tests pass
- Short summary of files changed + risks

Claude Code prefers CLAUDE.md. If both exist, put @AGENTS.md at the top of CLAUDE.md – or AGENTS.md stays a fallback and your shared rules never load. That’s the split-brain teams hit constantly.

Pro tip: Once: “Summarize the instructions you loaded.” If your rules aren’t quoted, fix filenames before any big edit.

3. Turn the sandbox on

Sandbox defaults to off in Claude Code (official sandboxing docs). Session command: /sandbox. Or user settings:

// ~/.claude/settings.json
{
 "sandbox": {
 "enabled": true,
 "allowUnsandboxedCommands": false
 }
}

Same spirit as the container/VM replies in the thread – without Kubernetes on afternoon one. Codex CLI’s stricter workspace-write default is the contrast people keep citing; Claude won’t protect you until you flip this.

4. Run the plan-before-code loop

  1. Feature + definition of done in chat.
  2. Require plan.md with paths and steps – no code yet.
  3. You annotate (cut scope, mark must-not-touch paths).
  4. Implement against the plan only after that.
  5. One-line append to changes.log when a chunk finishes.
  6. Fresh session or second model: “Review changes.log + diff; list bugs only.”

Same skeleton as the “research.md → plan.md → annotate N times” and “two agents, one log” patterns people rediscover every week. Fresh context on review catches more than one endless chat. Remember the thread bias? This loop is the part that actually transfers when you ignore the $500 flex.

5. Tier the work (without five subscriptions)

Inside one CLI, route by difficulty: lighter/faster model for renames and tests; strongest for architecture and security-sensitive paths. “Flash / pro / heavy / frontier” replies are the scaled version. Cheap version: say “use the fast model for this,” or split tickets so bulk never sits on the expensive tier.

Common pitfalls

  • Shared meters: Claude Code and claude.ai chat share the Pro/Max pool (Help Center). A long terminal day can throttle the web UI mid-afternoon – no separate meter. Settings → Usage.
  • Sandbox left off: Full shell until you enable it. Flip it before unattended runs.
  • Merging without a human gate: Draft + PR. Auto-merge ships confident nonsense.
  • Five parallel projects day one: Fun until your attention is the bottleneck. One repo, one definition of done.

After a week you can tell

Working setup: agent stops re-asking how to run tests, plan files match git, review passes find real bugs not style nits. People who simplified to one coding agent + a clear “what complete looks like” report less thrash than the swarm collectors.

Money: a lot of solid workflows sit on one $20 seat (Claude Pro or ChatGPT Plus) plus occasional API overflow. Multi-hundred-dollar stacks exist. They are not the median learning path.

Is the thrash from the tool – or from skipping plan files and burning the strong model on renames?

When not to use a full agentic setup

Skip the loop for one-line edits, pure “explain this file” learning, or writes where a wrong patch is catastrophic and you will not read the diff. Plain chat is enough. Skip multi-agent orchestration until single-agent + AGENTS.md + plan file feels boring – orchestration multiplies failure modes faster than output.

Mostly email triage or ops checklists? A narrow workflow tool beats a coding agent. Pick the tool that fits the job.

FAQ

Do I need Claude Max to start?

No. Pro includes Claude Code. Upgrade when usage walls hit, not before.

Claude Code vs OpenCode vs Codex – which first?

Already paying Anthropic? Claude Code. Want any model and a free MIT client (optional Go plan around $10/mo on their pricing roundups)? OpenCode. Living in ChatGPT Plus already (Plus commonly from $20/mo for CLI access; CLI itself Apache-2.0)? Codex. The CLI brand matters less than AGENTS.md, plan-before-code, and a review pass. Switch later – markdown instructions travel.

Why did my AGENTS.md rules get ignored?

Not the model being stubborn. With Claude Code, a CLAUDE.md or CLAUDE.local.md higher in the tree usually wins; AGENTS.md is fallback unless you @-import it or change project instruction settings (memory docs / v2.1.277+ behavior community guides call out). Ask which memory files loaded. Fix naming first. Keep the file under a couple hundred lines – bloated instruction files get partially dropped in practice.

Tonight: install one coding agent, drop a 30-line AGENTS.md in a real repo, enable sandbox, run one feature through plan.md → implement → changes.log → fresh review. Reread the Ask HN thread after – you’ll see the same skeleton under the expensive noise.

Related Tutorials