Autonomous Code Agent Patterns

Overview

The code-agent-patterns skill provides a standardized architectural framework for building, maintaining, and scaling autonomous Software Engineering (SWE) agents. It focuses on reliability, precision, and contextual awareness, drawing heavily from industry benchmarks like SWE-Bench and modern LLM-driven development tools.

When To Use

• Implementing autonomous code editing agents.
• Designing specialized agents for bug localization or patch generation.
• Orchestrating multi-agent systems for code review or testing.
• Building tools that need to interface with existing large codebases.

Core Patterns

1. Minimal Context Assembly

Avoid dumping the entire repository into the LLM context. Instead:

• Use dependency graph analysis to identify "impact zones."
• Implement "Breadth-First Context Discovery" (start with high-level symbols, drill down only when necessary).
• Support dynamic file inclusion based on the specific task (e.g., test files, related interfaces).

2. Symbol-Level Repository Indexing

Build a robust index that enables the agent to navigate the codebase as a human developer would:

• Symbol Maps: Index classes, functions, and interfaces along with their file locations.
• Call Graphs: Track function calls and dependencies to understand side effects.
• AST Analysis: Utilize Abstract Syntax Tree (AST) parsing to ensure structural awareness during code modifications.

3. Unified Diff & Search-Replace Patching

To minimize hallucination and merge conflicts:

• Prefer "Search-Replace" blocks over entire file rewrites.
• Validate that the "Search" block matches exact file content before applying the "Replace" content.
• Use unified diffs as a secondary representation for human review.

TDD Fix Loop Workflow

Agents must follow a strict "Red-Green-Refactor" loop for any bug fix:

1. Reproduce: Write a test case that captures the reported bug (assert failure).
2. Localize: Identify the source using logs, stack traces, or symbol-level search.
3. Patch: Apply minimal code changes to satisfy the failing test.
4. Verify: Run the test suite.
5. Rollback: If tests fail after patching, discard changes and restart the process.

Multi-Agent Coding Triad

To ensure high-quality output, structure teams as:

• Coder: Responsible for reading the codebase and proposing patches.
• Code Reviewer: Audits patches for logic errors, style violations, and architecture alignment.
• Test Engineer: Manages test execution and ensures sufficient coverage for the change.

Patch Validation Pipeline

Every proposed change must survive an automated gauntlet:

1. Syntax Check: Ensure the code is parsable.
2. Type Check: Validate against TypeScript/Python type definitions.
3. Unit Tests: Pass local unit tests.
4. Lint: Adhere to project linting standards.
5. Security Scan: Check against common patterns like OWASP Top 10.

SWE-Bench Lessons

• Mitigation Strategy: The most common failure mode is "blind editing" without understanding global constraints. Ensure agents have access to ADR (Architecture Decision Records).
• Hard Fixes: Avoid over-complex regex. Use AST-based transformations whenever possible.
• Context Drift: Periodically refresh context maps when making many changes in a single session.

References

• SWE-Bench
• KILO-KIT Documentation
• Aider: AI Pair Programming