Harness Router Adds Jev-Powered Tool Routing
Harness Router is an open-source decision layer that routes AI agents toward the right tool using Jev for fast ambiguity resolution and local MCTS for multi-step choices. It integrates through MCP, Codex and Claude Code hooks, or an explicit skill.
Harness Router targets a real weakness in tool-heavy agents: expensive models wasting reasoning on straightforward tool selection. Its architecture is promising, but the early benchmarks remain exploratory rather than proof of better end-to-end agent performance.
- –Codex and Claude Code hooks can intercept every pending tool call, allow good choices, and request replanning when a confident alternative exists.
- –Jev handles ordinary routing while MCTS explores downstream consequences locally without multiplying model calls.
- –Fail-open behavior and separate execution policies preserve sandboxing and authorization boundaries.
- –An exploratory 24-case benchmark reported 404.5 ms average router latency versus 3,842.1 ms for Codex selection, with identical choices in all cases.
- –The alpha project still needs held-out tests on genuinely ambiguous tool sets and full task completion, cost, and reliability metrics.
DISCOVERED
1h ago
2026-09-28
PUBLISHED
7h ago
2026-09-28
RELEVANCE
AUTHOR
Kamil Mosciszko