This article details the construction of an advanced agentic harness, moving beyond basic loops to incorporate structures for complex operations. Similar to real-world systems like air campaigns with mission planners and after-action reviews, production agents like Claude Code, Devin, and Cursor wrap a core agent loop with planning, execution, recovery, and verification mechanisms.

The core of this advanced harness is composition, building small, testable primitives: typed tools with Pydantic validation to prevent invalid arguments, a plan Directed Acyclic Graph (DAG) for parallel execution, tiered memory with retrieval budgets, a verification hierarchy, and multi-dimensional budgeting. These primitives address predictable failures in naive agents, such as hallucinated tool arguments or silent output propagation.

A city comparison agent serves as a running example, demonstrating how independent lookups can be parallelized within a DAG, while the final aggregation depends on their completion. The harness uses a pluggable LLMProvider interface, allowing for different backends and deterministic mock providers for testing. Typed tools, defined with Pydantic models, provide runtime validation and generate JSON schemas for tool-use APIs, ensuring that errors are caught before expensive tool calls. The plan is represented as a graph, validated before execution to prevent wasted resources on broken plans.