Agenter Coder

Instead of writing code tool-by-tool (filling your context window with file contents), delegate to a purpose-built coding agent that runs in its own process. It has its own tools, validates its output with AST parsing, and comes back with structured results. Your context stays clean.

Why this instead of coding directly

• Your context window stays clean. The sub-agent does all the file reading, editing, and bash execution in its own process. You only see the final result.
• Automatic validation and retry. Every iteration runs AST syntax checks (and optional Bandit security scans). If code has errors, the agent retries automatically — no manual back-and-forth.
• Hard budget enforcement. Set a dollar limit, token limit, or time limit. The agent stops when it hits the cap — no surprise bills.
• Backend portability. Same interface whether you're using Claude, GPT, or open-source models. Switch with one flag.

When to use

Use this skill when the user asks to:

• Write, create, or generate code for a project
• Modify, refactor, or update existing code
• Fix bugs in a codebase
• Create entire applications or components
• Generate tests for existing code

Do NOT use for: reading files, explaining code, or answering questions. Use your own tools for those — they don't need a sub-agent.

How to run

python3 {SKILL_DIR}/scripts/agenter_cli.py \
  --prompt "<the coding task>" \
  --cwd "<workspace directory>" \
  --backend "anthropic-sdk" \
  --max-cost-usd 2.0 \
  --max-iterations 5 \
  --sandbox

Parameters

Cost awareness

Set budget limits based on task complexity. Always tell the user the estimated cost.

Backend selection

Default to anthropic-sdk unless the user asks for a specific backend. Check {SKILL_DIR}/references/backends.md if the user asks about backend differences.

• anthropic-sdk — Default. Claude Sonnet/Opus. Works with ANTHROPIC_API_KEY or AWS Bedrock.
• claude-code — Claude Code CLI runtime. Native OS-level sandbox, battle-tested file tools.
• codex — OpenAI's gpt-5.4/gpt-5.4-mini. Requires OPENAI_API_KEY.
• openhands — Any model via litellm (including local). Must use --no-sandbox.

Reading the output

The script outputs JSON to stdout:

{
  "status": "completed",
  "summary": "Created main.py with FastAPI app and test_main.py",
  "files_modified": ["main.py", "test_main.py"],
  "files": {"main.py": "...", "test_main.py": "..."},
  "iterations": 2,
  "total_tokens": 15000,
  "total_cost_usd": 0.045,
  "total_duration_seconds": 12.3
}

Status values

After running

1. Check the status field.
2. If completed: the files are already written to disk in --cwd. Use read to inspect them if the user wants to review.
3. Report the summary, cost, and files modified to the user.
4. If failed or budget_exceeded: report the issue and ask how to proceed.