Agentic Workflows

The Agentic Loop

An agent operates in a cycle: Observe → Plan → Act → Verify. If you ask an agent to "Fix the memory leak in the bridge," it doesn't just guess. It reads the code, runs a memory profiler (like Valgrind or Instruments), identifies the leak, applies a fix, and runs the profiler again to verify the fix works. This is the difference between a chat bot and a software engineering agent.

// Agent Log:
// 1. Reading chess_bridge.cpp... identified missing 'delete' in chess_destroy.
// 2. Running 'make test'... tests passed, but leak confirmed via valgrind.
// 3. Applying fix to line 42...
// 4. Running 'make test' with valgrind... leak resolved. Task complete.

The tools you'll actually use

The AI agent space moves fast, but a few patterns are standard now:

• Claude Code, Cursor, Codex, Aider — terminal- or editor-integrated agents that read your repo, run your build, and iterate. Best for general feature work and large refactors.
• GitHub Copilot Workspaces / pull-request agents — task-scoped agents that take an issue and produce a PR. Best for narrow, well-specified bugs.
• The Anthropic and OpenAI SDKs directly — when you want to script your own agent for repository-specific work (e.g. "rename every Swift file's class prefix from PG to Game"). Cheaper than a generic agent for repeatable tasks.

What they share: tool use (read/write files, run commands), a planning loop, and the ability to keep going until the task succeeds or a budget runs out.

Automated Refactoring

Agents are perfect for large-scale, repetitive changes that are too complex for a simple search-and-replace. If you want to "Convert all raw pointers in the engine to std::unique_ptr," an agent can go through every file, update the types, fix the function signatures, and ensure the code still compiles at every step.

An agent is more than an AI; it is an AI with a toolbox and a mission. It closes the loop between thinking and doing.

The Human-in-the-Loop

The most effective agentic workflows are collaborative. You set the high-level strategy (the "Plan") and the agent handles the low-level execution (the "Act"). You act as the final validator, checking the agent's work and providing course corrections when it goes down a rabbit hole. This "Centaur" approach — human strategy plus AI execution — is how modern apps are built at 10x speed.

// User to Agent:
// "I want to add a new game: Sudoku. Research the rules, 
// design the C++ engine header, and show me the plan before 
// you start implementing."

What's next

We've mastered the development lifecycle with AI. Now, we need to get our apps into the hands of users. Next week, we'll cover Deploying: how to package your Mac and Windows apps, sign them for security, and distribute them to the world.

Week 51 is Deploying.