Harness Hub
Vol. I · No. 2Free edition

The Harness Daily

Morning editionharness-hub.org

The day's top story

OpenAI released GPT-6.2 Astra and Sol at lower API prices, and harnesses are already changing model catalogs and prompt caching so those models stay listed and are not rebilled on every step.

GPT-6.2 Astra and Sol, plus cache fixes

OpenAI released GPT-6.2 Astra and Sol with lower API prices; on the Artificial Analysis Intelligence Index, Astra ranks second behind Claude Fable 5.2. OpenCode now sends the system prompt as a leading developer message because GPT-6.2 does not cache one placed in the Responses instructions field, and jcode 1.0.1 keeps catalog updates that arrive while an agent is busy so models such as GPT-6.2 Sol are not dropped from the picker.

Sources: The Decoder: OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains; opencode v2.3.2; jcode v1.0.1

Key News in Harnesses

Copilot evals, review, models, and admin

GitHub put Copilot Evals in public preview, added MAI Code 2 and MAI Code 2 Flash, and expanded Fusion to more surfaces with new models and Terminal-Bench results. Code review can match depth to the pull request and detect secrets; the CLI can sign in with Microsoft Entra in preview, and Copilot agent activity can stream into Microsoft Agent 365.

Sources: GitHub Changelog: GitHub Copilot Evals is now in public preview; GitHub Changelog: MAI Code 2 and MAI Code 2 Flash in GitHub Copilot; GitHub Changelog: GitHub Copilot Fusion updates: more availability, benchmarks, and models; GitHub Changelog: Copilot code review: Adaptive effort, secret detection, and more

OpenCode shell timeouts and Browse rewrite

A shell command still running when its timeout expires now moves to the background instead of being killed, and a foreground timeout of 0 means the default 120 seconds. Browse 0.1.0 is a Rust rewrite that drops the Playwright dependency and exposes browser operations directly to the agent; the project reports 25x less memory and a 23x smaller build.

Sources: opencode v2.3.1

jcode 1.0 adds swarms and Cloud

jcode 1.0 plans swarm jobs in phases and parallel lanes, starts workers as dependencies finish, and shows the team's progress in one pinned view. jcode Cloud is included, with a way to move a session.

Sources: jcode v1.0.0

Qwen4-Max reaches OpenCode and Ante

Qwen4-Max is now available in OpenCode Go, and Ante's release notes list it among OpenRouter models. It is also a large Hacker News thread, separate from the GPT-6.2 launch.

Sources: Ante v0.3.6

The Market

Hottest tools · close of Nov 19, 2026

  1. T3 Code

    ▲ +3,005 stars in 2 days · 10 discussions

    T3 Code gained +3,005 stars in two days and 10 discussions as a control plane for running several coding agents from one interface.

  2. superpowers

    ▲ +2,976 stars in 2 days · 2 discussions

    superpowers gained +2,976 stars in two days as an agentic skills framework and software methodology.

  3. ECC

    ▲ +2,325 stars in 2 days

    ECC gained +2,325 stars in two days as a harness layer for skills, instincts, memory, and security.

  4. Ponytail

    ▲ +2,203 stars in 2 days

    Ponytail gained +2,203 stars in two days as a minimalism skill that pushes the agent to do less.

  5. last30days

    ▲ +2,159 stars in 2 days

    last30days gained +2,159 stars in two days as a skill that researches any topic across Reddit, X, YouTube, Hacker News, and Polymarket.

  6. AG-UI Protocol

    ▲ +2,051 stars in 2 days

    AG-UI Protocol gained +2,051 stars in two days as the protocol for connecting agents to frontends.

The Letters Page

What practitioners are complaining about this week

“Review is now the slow part”

Teams using coding agents hit this because Harvard researchers find those agents shifted the bottleneck to code review. Generating code is no longer the step that holds the work up; checking it is.

Sources: Hacker News · 316 points · 435 comments

“Codex capacity errors block sessions”

Codex users have been stopped by model-at-capacity errors, so a session cannot continue. The Codex team says that failure may now be fixed, which is a measure of how often it was getting in the way.

Sources: @thsottiaux on X · 6724 likes

“Compaction drops or fails context”

Antigravity users have been compacting often enough that Google is doubling the context window for Argon, Carbon, and Hydrogen and expects compaction per user turn to fall by 3.5x. Goose users hit a broken variant: Claude models refuse the compaction scratchpad, so Goose 1.60.1 drops it.

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.