How Claude Code Implements Harness Engineering for Robust AI Agents

Harness Engineering is presented as an industrial‑grade engineering framework for AI agents, aiming to solve four fatal problems that agents face in production: context memory loss, skill conflicts, semantic context clashes, and agent execution conflicts.

1. Bootstrap (Root State)

The bootstrap/ directory implements root‑state minimization, atomic state operations, and session‑level isolation. Critical fields such as session ID, project root, permission tags, and metrics are stored here, preventing global state pollution. Functions like switchSession and addCost are wrapped as atomic methods to avoid concurrent modifications.

2. Context (Memory Management)

The context/ module provides a four‑layer architecture (System, Project, Session, Turn) that isolates and compresses context. Automatic compression triggers when token thresholds are exceeded, preserving core decisions while discarding redundant tool results. Boundary markers ( compact_boundary) differentiate old and new history, and timestamps enable de‑duplication.

3. Skills (Capability Asset Library)

The skills/ directory standardizes capabilities as Skill objects. Each skill must declare a name, permission whitelist, allowed tools, execution mode, and associated files via BundledSkillDefinition. Permission whitelists ( allowedTools) block unauthorized tool usage, lazy loading avoids unnecessary execution, and an isEnabled switch allows dynamic activation or deactivation.

4. Coordinator + Tasks (Scheduling)

The coordinator/ and tasks/ modules form a traffic‑control system. Tasks follow a state machine ( pending→running→paused→completed/failed) to enforce ordered execution. Forked agents run in isolated contexts, preventing parent‑process pollution. Write tasks are serialized, read tasks allow limited concurrency, and resource locks guard against simultaneous file or resource modifications.

5. Query (Core Loop)

The query/ module acts as the brain, integrating all other modules. It enforces strict tool‑result pairing ( strictToolResultPairing), automatically assembles the four‑layer context, and validates stop reasons to avoid infinite loops. Self‑healing mechanisms retry, downgrade, or roll back on tool failures or conflicts.

Core Features and Iron Rules

Minimize global state and isolate sessions.

Layered, compressed, and bounded context management.

Standardize skills with permissions and boundaries.

Isolated, ordered scheduling with resource locking.

These principles together provide a full‑life‑cycle, controllable, and extensible agent framework, demonstrated by Claude Code's 510,000‑line codebase.