swe_agent

Software Engineering Agent

Autonomous coding agent demonstrating the agent + tools + memory pattern used by GitHub Copilot, Cursor, and Claude Code.

Core Concepts

Agent capabilities emerge from three components: • Tools: File operations, code execution, memory, metacognition • Prompts: Structured workflow encoding best practices • Memory: Persistent knowledge across sessions

Engineering Patterns

Five-phase workflow: Memory check → Planning → Execution → Learning → Completion Surgical edits: str_replace mode for precise changes Markdown tracking: Checkboxes in /memories/current_task.md Explicit evaluation: TaskStatusTool prevents premature termination

Quick Start

# Set credentials
export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export AZURE_OPENAI_API_KEY="your-key"

# Run example
cd examples/agents/swe_agent
python agent.py

What It Does

Runs three tasks showing memory usage:

Task 1: Create calculator module + tests Task 2: Add power function (reuses testing patterns from Task 1) Task 3: Add documentation (applies learned conventions)

Agent Architecture

File Tools: read_file, write_file (3 modes), list_directory, grep_search Execution: python_repl, bash_execute Memory: view, create, search, append, str_replace Metacognition: ThinkTool, TaskStatusTool

Memory structure:

agent_memory/
├── patterns/           # Reusable solutions
├── decisions/          # Dated decision logs
├── current_task.md     # Active plan with checkboxes
└── project_context.md  # High-level understanding

Configuration

Iteration limits (in Agent):

• Simple scripts: 10-20
• Multi-file projects: 30-50 (default)
• Complex refactoring: 50-100

Bash timeout (in create_coding_tools):

• 30s: Quick tests
• 60s: Test suites (default)
• 120s+: Large builds

Files

• agent.py - Main agent setup and example tasks
• scratch/agent_workspace/ - Generated code (isolated)
• scratch/agent_memory/ - Persistent memory

Key Insights

Tools are prerequisites, not guarantees. Quality depends on LLM capabilities, prompt guidance, and feedback from execution.

Prompts are software. They encode workflows, best practices, and completion criteria. Iterate and test them.

Memory enables learning. Patterns accumulate, mistakes are recorded, decisions are justified. Agent improves over time.

Completion needs explicit criteria. Clear requirements + TaskStatusTool + tests = reliable task completion.