Build a coding assistant agent with a search-read-edit-test workflow and tool creation capabilities.

Code Assistant

This example builds a coding assistant that follows a disciplined workflow: search first, edit minimally, test after every change, and iterate on failures. This is based on the coder agent included with supyagent.

The Agent YAML

agents/coder.yaml

name: coder
description: Expert programmer that builds, edits, debugs, and tests code
version: "1.0"
type: interactive

model:
  provider: anthropic/claude-sonnet-4-5-20250929
  temperature: 0.3        # Low temperature for precise code generation
  max_tokens: 8192         # Generous output for code blocks

system_prompt: |
  You are an expert programmer. You write clean, efficient code and follow a
  disciplined workflow to ensure quality.

  ## Workflow

  ### 1. Search First, Code Second
  Before making any changes:
  - Use search or find_files to locate relevant code
  - Use read_file to understand the existing implementation
  - Understand the patterns, naming conventions, and architecture before editing
  - Check for existing utilities or functions you can reuse

  ### 2. Edit Minimal
  Make the smallest change that solves the problem:
  - Use apply_patch for multi-file changes -- it is atomic and context-anchored
  - Use edit_replace for single surgical edits in one file
  - Never rewrite entire files unless creating something new
  - Match the existing code style -- indentation, naming, patterns

  ### 3. Test After Every Change
  After modifying code:
  - Run the relevant test suite with run_command
  - If tests fail, read the error output carefully
  - Fix the specific issue and rerun -- do not guess
  - If no tests exist, consider writing them

  ### 4. Iterate on Failure
  When something goes wrong:
  - Read the full error message -- stack traces, line numbers, error types
  - Fix the root cause, not the symptom
  - Do not retry the same failing approach -- try a different strategy
  - If stuck after 3 attempts, explain the issue clearly

  ### 5. Git Awareness
  Keep track of changes:
  - Use git_status to see what you have changed
  - Use git_diff to review changes before committing
  - Commit logical units of work with clear messages

  ## Tool Preferences
  - Prefer apply_patch over write_file for editing existing files
  - Prefer search over run_command with grep
  - Prefer read_file_lines over read_file for large files
  - Prefer edit_replace for simple single-point edits

  ## Scope Discipline
  - Only change what is needed for the task
  - Do not refactor surrounding code or fix unrelated issues
  - Do not add features that were not requested
  - Keep changes reviewable -- small and focused

tools:
  allow:
    - "*"

will_create_tools: true    # Can create new supypowers scripts at runtime

delegates: []              # Leaf agent -- does not delegate

limits:
  max_tool_calls_per_turn: 50    # Allow many tool calls for complex tasks

Key Design Decisions

Low Temperature

The temperature: 0.3 setting reduces randomness in code generation. This produces more consistent, predictable output -- important when editing existing code where precision matters.

`will_create_tools: true`

This flag instructs the agent on how to create new supypowers scripts when it encounters a task that existing tools cannot handle. The system prompt is automatically augmented with tool creation instructions.

No Delegates

The coder is a "leaf" agent in the delegation hierarchy. It does the actual implementation work rather than delegating further. In a multi-agent setup, the planner delegates to the coder.

High Tool Call Limit

max_tool_calls_per_turn: 50 allows the coder to perform complex multi-step operations (search, read, edit, test, fix) within a single turn.

Using the Coder

Interactive Mode

supyagent chat coder

You: Add input validation to the User model in src/models/user.py

coder> Let me search for the User model first.
  [tool: search__search_files] pattern="class User" ...
  [tool: files__read_file] path="src/models/user.py" ...

I can see the User model. Let me add email validation and
length constraints.

  [tool: edit__edit_replace] ...
  [tool: shell__run_command] command="pytest tests/test_user.py" ...

Done. I added email format validation using a regex pattern and
length constraints for the name field. All tests pass.

As a Delegate

The coder works best when called by a planning agent:

agents/planner.yaml

name: planner
description: Plans and orchestrates complex tasks
delegates:
  - coder

supyagent plan "Add a REST API endpoint for user registration"

The planner breaks down the task, then delegates implementation steps to the coder.

Customization

Language-Specific Coder

agents/python-coder.yaml

name: python-coder
model:
  provider: anthropic/claude-sonnet-4-5-20250929
  temperature: 0.3
system_prompt: |
  You are a Python expert. You follow PEP 8, use type hints,
  and write pytest-compatible tests.

  Always use ruff for linting and formatting.
  Prefer pathlib over os.path.
  Use pydantic for data validation.
tools:
  allow: ["*"]

Read-Only Code Reviewer

agents/reviewer.yaml

name: reviewer
model:
  provider: anthropic/claude-sonnet-4-5-20250929
  temperature: 0.5
system_prompt: |
  You are a code reviewer. Read the code, identify issues,
  and suggest improvements. Do NOT make changes yourself.
tools:
  allow:
    - "files__read_*"
    - "search__*"
    - "find__*"
    - "git__*"
  deny:
    - "files__write_*"
    - "edit__*"
    - "shell__*"

Building Agents -- Agent configuration reference
Multi-Agent Systems -- Delegation patterns
Custom Tool -- Creating new tools

Code Assistant

On this page