Harness Hub
Vol. I · No. 1Free edition

The Harness Daily

Morning editionharness-hub.org

The day's top story

New models (GPT-6.2, Qwen4 Max, Gemini 4) reached harnesses within a day, while Copilot added evals, review tooling and enterprise controls.

GitHub Copilot adds evals, adaptive review and enterprise controls

Copilot Evals is in public preview, and Copilot code review now adapts its depth to each pull request. Enterprises can stream Copilot agent activity into Microsoft Agent 365, and Copilot CLI has Entra sign-in in preview. MAI Code 2 models were added, and some older models were deprecated on Nov 16.

Sources: GitHub Changelog: GitHub Copilot Evals is now in public preview; GitHub Changelog: Copilot code review: Adaptive effort, secret detection, and more; GitHub Changelog: Monitor GitHub Copilot activity through Microsoft Agent 365; GitHub Changelog: GitHub Copilot CLI adds Microsoft Entra sign-in as a preview

Key News in Harnesses

OpenCode 2 rollout, background shell commands, Browse in Rust

OpenCode 1 now offers an in-app upgrade to OpenCode 2.0. Plugins written for OpenCode 1 need an update. In v2.3.1, shell commands that hit their timeout move to the background instead of being killed. OpenCode Browse 0.1.0 was rewritten in Rust, and its browser tools are available in codemode.

Sources: opencode v1.19.6; opencode v2.3.1

Swarm-style harnesses: jcode 1.0 and Missions

jcode v1.0.0 reworks swarms to plan work in phases and parallel lanes, and launches jcode Cloud. Separately, @amasad announced Missions, an agentic swarm harness that works toward a goal within a set budget.

Sources: jcode v1.0.0

Claude Code Action: breaking change on AGENTS.md in PRs

v1.0.277 now uses the base branch's AGENTS.md on pull requests, flagged as a breaking change. Later releases use the default branch's config for PRs from authors without write access.

Sources: Claude Code Action v1.0.277; Claude Code Action v1.0.279

The Market

Hottest tools · close of Nov 18, 2026

  1. T3 Code

    ▲ +1,889 stars in 2 days · 10 discussions

    An open-source control plane for running several coding agents from one interface, up about 1,900 stars in two days.

  2. DeepSeek Harness

    ▲ +847 stars in 2 days · 8 discussions

    DeepSeek's open agent harness, built around plugins, with 8 discussions and a Reddit thread about building an air-gapped RAG assistant on it.

  3. AG-UI Protocol

    ▲ +1,611 stars in 2 days

    The AG-UI protocol for connecting agents to frontends gained stars and shipped a new npm release on Nov 17.

  4. opencode

    ▲ 20 discussions

    OpenCode has the OpenCode 2 upgrade, a Rust rewrite of OpenCode Browse, and Qwen4 Max support all in the same news cycle.

  5. Open Code Review

    ▲ +1,053 stars in 2 days

    Alibaba's code review CLI mixes deterministic checks with an agent, which fits the review-bottleneck discussion; it shipped v1.13.11 with branch filtering for sessions.

  6. Claude Code

    ▲ 83 discussions

    Claude Code has the most discussion of any tool today (83), including the team's post on verification and new MCP and GitHub Action changes.

The Letters Page

What practitioners are complaining about this week

“Agents holding credentials and running untrusted code”

Anyone running coding agents locally has to decide what secrets and file access the agent gets. People are describing external sandboxes (WebAssembly, QEMU, Apple Container) and ways to run agents without handing over credentials. Daytona now offers macOS sandboxes.

Sources: @simonw on X · 1148 likes; @ivanburazin on X · 680 likes

“Wasted tokens and unclear cost”

Users who give agents open-ended goals see them misunderstand the task and burn tokens. Tools like Tokendoctor try to show where Claude Code usage overpays. One comment argues that a sprawl of specialized subagents costs more than a single agent.

Sources: @amasad on X · 659 likes; @emollick on X · 1737 likes

“Context loss and compaction”

Long sessions lose detail when history is compacted. Antigravity says doubled context windows cut compaction per user turn by 3.5x. Agno added an endpoint to list a session's compactions with token counts, and Hive changed when memory notes reach the agent.

Sources: @_mohansolo on X · 5505 likes

In Other News

Situations Vacant

New roles at harness companies · all open roles · the job market

Classifieds

Front page, letters and market written by AI (grok-4.7) from the releases, news, Hacker News, Reddit and X threads of the last two to three days and GitHub star growth; every claim links to its source. Other news, vacancies and classifieds come straight from the site's own listings.

Download share card

The Harness Hub newsletter

New harness companies, projects, releases and the monthly State of Harness report, in your inbox.