Releases·Tool Release
I built a tool router for AI coding agents that can run 4,096 MCTS simulations before choosing an action
A developer released harness-router, an open-source MIT-licensed MCP server that routes tool calls for agentic coding harnesses. It keeps obvious actions cheap, uses a small decision model (Jev) for ambiguous choices, and runs Monte Carlo Tree Search when multi-step consequences matter. The author reports exploratory benchmark results against Codex tool selection, though the MCTS budget was tuned on the same test cases.
3 engagement1 sourceSat, Sep 26, 2026, 08:40 PM
AI Overview
- harness-router is an MIT-licensed Python MCP server using Jev to route obvious tool calls cheaply and 4,096-simulation depth-3 MCTS for ambiguous ones
- Dev benchmarked it against standard Codex tool selection, calling results exploratory since the MCTS budget was tuned on the same test cases
- Its architecture: LLM for open-ended reasoning, small decision model for closed choices, MCTS search when future actions matter
Entities
Related
Products·building
Harness Router — decision layer that routes tool calls for AI agent harnesses
Harness Router is an open-source, framework-agnostic tool router that inserts a decision layer before an LLM agent executes tool calls, instead of letting the model pick blindly from a large tool set. It hooks into agent harnesses via Codex hooks (SessionStart to discover tools, PreToolUse to vet each call) and can allow, re-plan, or fail open. It targets developers building agentic coding systems who need more reliable tool selection.
Harness Router is an open-source, framework-agnostic tool router that inserts a decision layer before an LLM agent executes tool calls, instead of letting the model pick blindly from a large tool set. It hooks into agent harnesses via Codex hooks (SessionStart to discover tools, PreToolUse to vet each call) and can allow, re-plan, or fail open. It targets developers building agentic coding systems who need more reliable tool selection.
1 engagement1 sourceredditMon, Sep 28, 2026, 07:36 AM
Products·launched
Agent framework where the agent is a single file
A lightweight agent framework that defines the entire agent logic in one file. It uses LLMs to power autonomous decision-making and tool use, simplifying development for builders who want minimal boilerplate.
A lightweight agent framework that defines the entire agent logic in one file. It uses LLMs to power autonomous decision-making and tool use, simplifying development for builders who want minimal boilerplate.
2 engagement1 sourcehackernewsMon, Jul 13, 2026, 10:14 AM
Releases·Tool Release
Aidress registry integrates with LangChain for cross-boundary agent discovery
Aidress, an open zero-commission registry for agent discovery and trust, has integrated with LangChain, enabling agents to dynamically discover capability schemas and verify counterparty identity across organizational boundaries. The integration addresses a key limitation in multi-agent systems where agents rely on hardcoded local tools.
Aidress, an open zero-commission registry for agent discovery and trust, has integrated with LangChain, enabling agents to dynamically discover capability schemas and verify counterparty identity across organizational boundaries. The integration addresses a key limitation in multi-agent systems where agents rely on hardcoded local tools.
1 engagement1 sourceredditSat, Aug 8, 2026, 03:52 AM
Social·Community
Redditor tests Laya as local model router for coding agents; zero of 48 routing decisions cleared the confidence threshold
A developer built a local proxy that routes coding-agent turns between DeepSeek Flash and Pro using Laya, a small open-source decision model, as the tier classifier. Across 48 hand-labelled prompts, Laya never reached the 0.8 confidence threshold needed to switch tiers, mislabeling 10 of 14 hard prompts as 'balanced'. The router stays in shadow mode while the author compares routed decisions against real outcomes, and a small test found Flash outscoring Pro.
A developer built a local proxy that routes coding-agent turns between DeepSeek Flash and Pro using Laya, a small open-source decision model, as the tier classifier. Across 48 hand-labelled prompts, Laya never reached the 0.8 confidence threshold needed to switch tiers, mislabeling 10 of 14 hard prompts as 'balanced'. The router stays in shadow mode while the author compares routed decisions against real outcomes, and a small test found Flash outscoring Pro.
4 engagement1 sourceredditThu, Oct 15, 2026, 03:35 PM
Products·building
MARGINAL: Open-source runtime governor for AI coding agents
MARGINAL is an open-source runtime governor that sits in the loop of AI coding agents to evaluate whether the next action is worth executing, preventing wasteful compute. It starts in observation mode, collects evidence, and only enforces restrictions when patterns are strong, failing open otherwise. It solves the problem of agents wasting tokens on repetitive or low-value actions, benefiting developers using AI coding agents.
MARGINAL is an open-source runtime governor that sits in the loop of AI coding agents to evaluate whether the next action is worth executing, preventing wasteful compute. It starts in observation mode, collects evidence, and only enforces restrictions when patterns are strong, failing open otherwise. It solves the problem of agents wasting tokens on repetitive or low-value actions, benefiting developers using AI coding agents.
1 engagement1 sourceredditTue, Aug 18, 2026, 07:30 AM