← Voltar aos guias

Claude Code Agent Blueprint: Build, Test, Control Costs

Decide what your coding agent should do

Before configuring Claude Code, define one repeatable job. A focused agent is easier to secure, test, and budget than a general-purpose assistant. Good starting projects include reviewing pull requests, fixing failing tests, modernizing a small module, generating API clients, or preparing release notes.

Write the agent's input, allowed actions, and success condition in plain language. For example: “Inspect the failing checkout tests, identify the smallest safe fix, run the relevant test suite, and summarize the changed files.” This becomes the basis for your project instructions and evaluation checklist.

Set up Claude Code with a compatible API

Install Claude Code using the current method in its official documentation, then configure authentication through environment variables rather than placing a key in a repository. With a compatible relay such as 59API, set the API base URL to https://api.59api.com and provide the token variable required by your Claude Code version and 59API account.

A typical shell setup resembles: export ANTHROPIC_BASE_URL=https://api.59api.com followed by export ANTHROPIC_API_KEY=your_token. Confirm the exact variable names and any required model setting in the current Claude Code and 59API documentation, then run Claude Code from the project directory. Never commit these values to .env files that are tracked by Git.

59API is useful when you want pay-as-you-go access without committing to a large subscription. It supports Claude Opus, Sonnet, Haiku, and Fable models, as well as GPT models, while remaining compatible with Claude Code, Codex, and OpenAI SDK workflows. Its native official-quality model access avoids the uncertainty of a downgraded substitute, and its low relay pricing can make experimentation more affordable.

Give the agent a controlled workspace

Create a project instruction file that explains the stack, commands, coding standards, and boundaries. Tell the agent which directories it may change, which tests must pass, and whether it may install packages or access the network. Keep the instructions short and operational.

• Describe the package manager, runtime version, test command, formatter, and linter.
• Require a plan before broad edits and a summary after each task.
• Tell the agent not to modify secrets, deployment configuration, migrations, or generated files without approval.
• Ask it to inspect existing patterns before introducing a new abstraction.

Use Claude Code's permission prompts and sandboxing features where available. Start with read-only analysis, approve individual writes, and expand access only after the workflow behaves predictably. Human approval should remain required for destructive commands, production changes, database operations, and dependency upgrades.

Choose the right model for each stage

Do not use the most capable model for every request. Use Haiku for quick file classification, simple explanations, and repetitive edits. Use Sonnet for everyday implementation, debugging, and test generation. Reserve Opus for difficult architectural reasoning, multi-file refactoring, or ambiguous failures. If Fable is available in your 59API account, compare it against your task's quality and latency requirements rather than assuming one model always wins.

A practical pattern is to use a fast model to collect context, a stronger model to make the decision, and a fast model again to format the final report. Check model identifiers and pricing before automating this routing.

Test the agent like software

Create a small evaluation set of real tasks with known outcomes. Include a bug fix, a feature request, a refactor, and a task where the correct behavior is to ask for clarification. Measure whether the agent changes only expected files, passes tests, follows project conventions, and explains uncertainty accurately.

Run the agent in a disposable branch or container. Capture prompts, model names, token usage, changed files, test results, and approval decisions. This makes regressions visible and helps you calculate the cost per completed task.

Simple launch checklist

• Define one narrow coding workflow and its success criteria.
• Configure Claude Code with a secured 59API token and the correct base URL.
• Add project instructions, test commands, and explicit file boundaries.
• Start with manual approvals and least-privilege access.
• Select Haiku, Sonnet, Opus, or Fable according to task difficulty and budget.
• Evaluate the agent on real tasks in an isolated branch or container.
• Track usage, failure rate, review time, and cost per successful task.

If you want to try Claude Code without a large upfront commitment, you can sign up for 59API, use its pay-as-you-go relay, and check whether its referral rebate improves your effective cost as your usage grows.