Releases·Tool Release
Show HN: FrontierHarness eval runs one model across 9 harnesses — cost per pass varies 17x
A Show HN post introduces FrontierHarness, an evaluation testing 9 agent harnesses (including Claude Code, Codex, Pi, Exo, and DeepSeek Harness) on 30 tasks with the same model. Pass rates ranged from 50% to 67% while cost per pass varied 17x, showing harness choice now rivals model choice in driving agent performance and economics.
164 engagement3 sourcesWed, Sep 2, 2026, 04:14 PM
Developments
Follow-up HN discussion notes Claude Code cost 5.6x as much for the same pass rate as cheaper harnesses.
Wed, Sep 9, 2026, 07:11 PM
AI Overview
- FrontierHarness benchmark runs the same model through 9 agent harnesses and finds cost per pass varies 17x
- Source published at frontierharness.org and shared on Hacker News on 2026-09-02; a Reddit thread followed on 2026-09-09
- A 2026-09-09 Hacker News discussion of FrontierHarness reports Claude Code cost 5.6x more for the same pass rate
- FrontierHarness's 360 runs on 30 tasks consumed 2 billion tokens; Claude Code and DSH Creator both passed 19 tasks
- Pass rates ranged 50% to 67%; median cost per pass ranged $1.05 to $18.34, with retries and manual cleanup counted toward total spend
- FrontierHarness recommends Codex for default use, Pi for high-volume repeat jobs, Exo when retries are cheap, DSH for wall-clock speed
Related
Products·launched
Oqoqo: Build realistic evals and custom benchmarks for agent tasks
Oqoqo is a platform for creating realistic evals and custom benchmarks for real-world tasks, measuring how agent-friendly product surfaces are against various AI coding agents. It helps developers regression test MCP, CLI, skills, SDK, and other agent-facing interfaces, and allows creating and sharing custom benchmarks.
Oqoqo is a platform for creating realistic evals and custom benchmarks for real-world tasks, measuring how agent-friendly product surfaces are against various AI coding agents. It helps developers regression test MCP, CLI, skills, SDK, and other agent-facing interfaces, and allows creating and sharing custom benchmarks.
1 engagement1 sourcehackernewsMon, Aug 10, 2026, 09:27 PM
Products·launched
Multiplex: SSH/tmux/herdr terminal for Vision Pro and iPad
Multiplex is an open-source terminal app for visionOS and iPadOS that provides a multi-window SSH/tmux experience with first-class tmux integration, each session in its own window. It includes an Agent key bar, file viewer, and browser to support using Claude Code, Codex, and Pi for code review, and recently added support for herdr sessions. It solves the problem of needing a mobile terminal with robust tmux support and AI coding tool integration.
Multiplex is an open-source terminal app for visionOS and iPadOS that provides a multi-window SSH/tmux experience with first-class tmux integration, each session in its own window. It includes an Agent key bar, file viewer, and browser to support using Claude Code, Codex, and Pi for code review, and recently added support for herdr sessions. It solves the problem of needing a mobile terminal with robust tmux support and AI coding tool integration.
1 engagement1 sourcehackernewsThu, Aug 13, 2026, 12:03 PM
Releases·Tool Release
dsh-plugin-dev-skills: Agent Skill standardizes DeepSeek Harness plugin development
A new Agent Skills skill, dsh-plugin-dev-skills, codifies the conventions for developing DeepSeek Harness (DSH) plugins into an executable standard. It works across Claude Code, Codex, DSH, VS Code Copilot, and any compatible agent, consolidating official documentation into a single skill.
A new Agent Skills skill, dsh-plugin-dev-skills, codifies the conventions for developing DeepSeek Harness (DSH) plugins into an executable standard. It works across Claude Code, Codex, DSH, VS Code Copilot, and any compatible agent, consolidating official documentation into a single skill.
35 engagement1 sourcegithubSat, Aug 15, 2026, 08:14 AM
Products·building
Local-first agent optimized for fast prompt processing
A lightweight AI agent designed for local LLM inference, prioritizing fast prompt processing over cloud-optimized features. It solves the problem of slow context compression and startup times when using local models, targeting users who prefer privacy and speed.
A lightweight AI agent designed for local LLM inference, prioritizing fast prompt processing over cloud-optimized features. It solves the problem of slow context compression and startup times when using local models, targeting users who prefer privacy and speed.
0 engagement1 sourceredditThu, Aug 20, 2026, 04:48 PM
Products·building
Ormah — portable long-term memory for AI coding agents
Ormah shares long-term memory across AI coding agents like Claude Code, Codex and Pi, injecting relevant context before each agent response. It stores memory locally as readable Markdown, links related memories, merges duplicates, and expires stale ones. It solves re-explaining preferences and decisions to each new agent, for developers using multiple AI agents.
Ormah shares long-term memory across AI coding agents like Claude Code, Codex and Pi, injecting relevant context before each agent response. It stores memory locally as readable Markdown, links related memories, merges duplicates, and expires stale ones. It solves re-explaining preferences and decisions to each new agent, for developers using multiple AI agents.
1 engagement1 sourceredditSat, Sep 5, 2026, 09:19 PM