Agentic engineering

Agent engineering is a disciplined approach to software development where professionals use Large Language Models (LLMs) to accelerate their work while remaining accountable for the code produced. It represents the professional end of specturm compared to “vibe coding” which is often fast and prompt-driven without deep attention to the underlying code.

But how does it work?

Agentic eneginering works through several fundamental core patterns designed to overcome LLM limitations and maximise developer leverage.

This includes:

• Context management via subagents: Because LLMs have finite context limits (working memory), agentic engineering uses subagents to handle specific tasks. This prevents the main agent from becoming overwhelmed by too much informration.

• Task decomposition: Complex projects are broken down into smaller, managable tasks suitable for a “swarm” of parallel agents. For example, a planner agent might create a list of tasks, which are then distributed to multiple “worker” agents. In extreme cases, these swarms can involved hundred of agents writing millions of lines of code.

• Parallel execution: Engineers often run multiple agent sessions simultanesouly - sometimes in a terminal grid or across different git worktrees - to work on different features at once.

• Tool integration: Agents are equipped with a “harness” of tools that allow them to edit code, perform semantic searches, grep strings, and run terminal commands.

• Verification: The process often concludes with a “judge agent” or automated test suites to verify that the work meets the required specifications