DEV Community

brian austin
brian austin

Posted on

Claude Code agents: the .claude/ settings that actually make them faster

#claudecode #ai #agents #productivity

Claude Code agents: the .claude/ settings that actually make them faster

Everyone's debating whether to go hard on agents or keep things simple. Meanwhile, the .claude/ folder has been sitting in your repo this whole time, quietly determining how fast (or slow) your agents actually run.

Here's what I learned after running Claude Code agents in production: the bottleneck is rarely the model — it's the configuration.

The two things everyone gets wrong

The "go hard on agents" camp says: spin up parallel agents, let them collaborate, emergent intelligence wins.

The "keep it simple" camp says: one agent, clear task, don't over-engineer.

Both camps are ignoring the third option: configure your agents properly before you decide which camp you're in.

What's actually in your .claude/ folder

If you haven't opened it recently:

.claude/
├── settings.json      ← this is where performance lives
├── CLAUDE.md          ← project context (agents read this every run)
└── commands/          ← custom slash commands
    └── deploy.md
Enter fullscreen mode Exit fullscreen mode

The settings.json file is doing a lot of quiet work.

The settings that change agent speed

1. apiBaseUrl — the one most people miss 

{
  "apiBaseUrl": "https://simplylouie.com/api"
}
Enter fullscreen mode Exit fullscreen mode

By default, Claude Code hits the Anthropic API directly at pay-per-token rates. If you're running agents that make 50+ API calls per session (totally normal for complex tasks), those costs compound fast.

Changing apiBaseUrl to a cheaper endpoint doesn't slow the model down — you're still getting Claude Sonnet or Haiku, just routed differently. SimplyLouie's developer API costs $2/month flat, which for high-frequency agent workflows is dramatically cheaper than per-token billing.

2. CLAUDE.md length — the hidden latency source

Every time an agent starts a new task, it reads your entire CLAUDE.md. I've seen CLAUDE.md files over 4,000 tokens. That's 4,000 tokens of context window used before the agent does anything.

Keep CLAUDE.md under 500 tokens. Put architecture docs somewhere the agent can fetch them on-demand instead:

# Project: MyApp
Stack: Node.js + PostgreSQL + Redis
Deploy: `npm run deploy` — runs tests first
Docs: See /docs/ folder for architecture details
Enter fullscreen mode Exit fullscreen mode

That's it. 50 tokens. The agent finds what it needs when it needs it.

3. Custom commands — where agents actually get fast

The .claude/commands/ folder is underused. Instead of describing a complex multi-step task every time:

<!-- .claude/commands/ship.md -->
Run tests, fix any failures, commit with conventional commit message, push to main, confirm CI passes.
Enter fullscreen mode Exit fullscreen mode

Now your agent workflow is:

/ship
Enter fullscreen mode Exit fullscreen mode

One command. The agent knows exactly what "ship" means in your project context. No ambiguity, no re-explanation.

The multi-agent pattern that actually works

Here's the configuration for running parallel Claude Code agents without them stepping on each other:

{
  "apiBaseUrl": "https://simplylouie.com/api",
  "permissions": {
    "allow": [
      "Bash(git:*)",
      "Bash(npm:*)",
      "Read(**)",
      "Write(src/**)"
    ],
    "deny": [
      "Bash(rm -rf:*)",
      "Write(*.env)"
    ]
  }
}
Enter fullscreen mode Exit fullscreen mode

The deny rules prevent agents from accidentally nuking your environment. The Write(src/**) restriction means agents can only write to source files — no touching config, no touching .env.

With this setup, you can run one agent on frontend, one on backend, one on tests — they share the same codebase but can't interfere with each other's scope.

The benchmark nobody talks about

The HN debate about whether agents are fast or slow is missing the baseline: an unconfigured agent doing 20 API calls at $0.003/call = $0.06 per session. At 100 sessions/month = $6/month just for one developer.

With a flat-rate API and a lean CLAUDE.md: same 20 calls, same capability, $2/month total. The difference isn't the model. It's the configuration.

What this means for "go hard on agents"

If you've been hesitant to go hard on agents because of cost or unpredictability, the .claude/ settings are your answer:

  1. apiBaseUrl → cap your cost at a flat rate
  2. Short CLAUDE.md → faster context loading
  3. Custom commands → encode your workflow patterns
  4. Permission rules → prevent agents from doing damage

Go hard on agents. Configure them first.

Running Claude Code agents and want to cut the API costs? SimplyLouie's developer API is $2/month flat — same Claude models, one base_url change in your .claude/settings.json.

Top comments (0)