Harness Router

Product Introduction

1. Definition: Harness Router is a specialized decision layer and routing engine designed for AI coding agents and developer harnesses. It operates as a Model Context Protocol (MCP) server, functioning as middleware that intercepts and intelligently selects the most appropriate tool for an agent to use before execution.
2. Core Value Proposition: It exists to eliminate wasteful AI model reasoning on trivial tool choices and to provide a cost-effective, high-accuracy decision-making layer for complex, multi-step workflows. Its primary value is in reducing AI agent latency, cutting operational costs, and improving decision reliability in automated coding and development tasks.

Main Features

1. Dual-Mode Operation (Skill & Hook): Harness Router can be used explicitly as a skill that an AI agent invokes to offload a tool-selection decision. More powerfully, it can be configured as a PreToolUse hook within the Codex editor, automatically intercepting every tool call for routing without requiring agent intervention. This provides both flexibility and enforced optimization.
2. Tiered Decision Engine: The system employs a smart, multi-tiered routing strategy. Simple, obvious choices are handled by a fast route function. For ambiguous decisions, it escalates to use Jev, a cost-efficient structured prediction model. For scenarios where downstream consequences matter, it can deploy a bounded Monte Carlo Tree Search (MCTS) algorithm (route_mcts) to simulate future states and select the optimal first action.
3. Native MCP Integration & State Management: It runs as a lightweight, persistent stdio MCP server, not a heavy helper process. It efficiently discovers and caches the full tool inventory from Codex on startup, then uses a compact state representation (goal, observation, tool shortlist) to make rapid decisions, minimizing overhead and connection latency.

Problems Solved

1. Pain Point: Inefficient AI Agent Spending. Large Language Models (LLMs) like Claude Sonnet or GPT-4 often spend significant reasoning time and tokens deliberating over simple, closed-choice tool selections (e.g., read_file vs. write_file), which is computationally expensive and slow.
2. Target Audience: Developers and engineers building or operating AI-powered coding assistants (e.g., using Codex, Cursor, or similar agentic IDEs); teams seeking to optimize the cost and performance of autonomous coding workflows; MCP server developers who want to add intelligent routing to their tool suites.
3. Use Cases: Automated Code Refactoring: Deciding whether to read, edit, or run_tests on a file. Multi-Step Debugging: Routing an agent through a sequence of inspect_logs, search_docs, and apply_fix tools. Build & Deployment Automation: Choosing between lint, build, deploy_staging, and run_integration_tests actions based on current state.

Unique Advantages

1. Differentiation: Unlike manual tool-calling or relying solely on the agent's native reasoning, Harness Router inserts a dedicated, optimized decision layer. Compared to using a more expensive LLM for all decisions, it leverages the cost-effective Jev model for commensurate accuracy at ~1/300th the cost, as validated by third-party benchmarks.
2. Key Innovation: The adaptive decision pipeline that combines rule-based fast routing, model-based classification (Jev), and algorithmic search (MCTS) based on decision complexity. This "fail-open" architecture ensures robust performance without breaking existing workflows, while the side-effect-free MCTS simulation allows for consequence-aware planning without executing dangerous actions.

Frequently Asked Questions (FAQ)

1. How does Harness Router reduce AI coding agent costs? It offloads simple tool-selection decisions from expensive primary LLMs (like Claude Sonnet 5) to the far more cost-efficient Jev model, which shows equivalent accuracy on workflow benchmarks for a fraction of the cost, leading to significant savings per task.
2. What is the performance impact of using Harness Router on tool call latency? Benchmark data shows it can reduce decision latency by approximately 9-12x. For ordinary route calls, mean latency dropped from ~3842ms with Codex alone to ~405ms. For deeper route_mcts searches, latency was reduced from ~4772ms to ~388ms.
3. Is Harness Router compatible with my existing MCP tools and Codex setup? Yes, it is designed as a drop-in MCP server. After installation and configuration, it automatically discovers your existing MCP tool inventory and integrates with Codex via the PreToolUse hook or skill call, requiring no changes to your individual tool servers.
4. When should I use the MCTS (route_mcts) feature versus the standard route? Use the standard route for most single-step, obvious tool choices. Escalate to route_mcts when the decision has significant multi-step consequences and ambiguity, such as planning a sequence of debugging actions or refactoring steps, where simulating 3-4 steps ahead improves the final outcome.
5. What happens if Harness Router is uncertain or fails to make a decision? The system is designed to fail open. If the router is uncertain, times out, or encounters an error, the original tool call proposed by the AI agent proceeds normally. This ensures reliability and does not block the agent's workflow, maintaining sandbox and approval rules.