HarnessRouter Open-Sources a Universal Protocol for Agent Harnesses

HarnessRouter Open-Sources a Universal Protocol for Agent Harnesses

A YC-backed project called HarnessRouter released UHP (Unified Harness Protocol) under Apache 2.0, giving developers one API to swap between Claude Code, Codex, Gemini CLI, and over 20 other agent harnesses [4]. It's compatible with the OpenAI Responses API, ships Docker-deployable, and supports persistent sessions, streaming, file handling, and cancellation — no hosted middleman required if you self-host [5][6].

Alibaba's Qoder has already adopted it, which is a meaningful early signal — this isn't just a hobby spec. The pitch is straightforward: stop building your product around one vendor's harness quirks, and let the protocol handle switching costs when a better model or cheaper harness shows up next month.

This matters more than it looks on the surface. Every "unification layer" in tech history has either become critical infrastructure or died quietly, and the difference is usually adoption velocity in the first 90 days. Worth watching whether this becomes the de facto standard or just one of several competing specs — but the vendor lock-in problem it's solving is real and growing as the harness landscape fragments.

Kimi K3.1 Pro Closes the Gap on Claude Opus — and It's Going Open Weight

Moonshot AI dropped Kimi K3.1 Pro on October 23 — a 2.8T parameter MoE model (104B active) with a 1M token context window, scoring 55 on the Artificial Analysis Intelligence Index, just three points behind Claude Opus 5.5 [7][8]. It's strong on coding (Terminal-Bench around 65-67%), handles 3D modeling, and reportedly hallucinates less than prior Kimi releases [9]. A "Highspeed" variant runs 2-3x faster, and API pricing undercuts the frontier labs at $2/$10 per million tokens [8].

Full open weights land November 4. Notably, Moonshot deliberately dialed back cyber capabilities ahead of the open release — a small but telling acknowledgment that open-weighting a near-frontier model carries real misuse considerations [7].

The open-vs-closed gap keeps shrinking, and three points on a benchmark is now the margin, not a canyon. For any team evaluating build-vs-buy on model infrastructure, this is the kind of release that should recalibrate your cost assumptions every quarter, not just once a year.

Net Forge: A Game Engine Built for Agents, Not Humans

Net Forge, an MIT-licensed TypeScript game engine, launched publicly October 23-24 with a clear thesis: stop retrofitting human-first tools for AI agents, build the tool for the agent from scratch [10][11]. It combines a shared ECS with Rapier physics for browser and server, three.js/WebGPU rendering, native MCP/CLI integration, and ten ready-made multiplayer templates [10].

The key design choice: agents build and edit games through a single typed src/game.ts file, with bot playtesting baked in for validation. That's a deliberate constraint — fewer files, fewer places for an agent to lose context, and a built-in feedback loop so the agent can check its own work [11].

This is a small release, but it's part of a pattern worth tracking: infrastructure increasingly being designed agent-first, human-second. Expect more of this across domains in the next year — tools that assume their primary user doesn't sleep, doesn't get confused by sprawling file trees, and needs tight validation loops instead of a debugger.

Sweden Blocks ChatGPT and Claude on Government Computers

Starting today, October 26, Sweden's Government Offices — including the Prime Minister's Office and ministries — blocked browser access to ChatGPT and Claude on staff machines, following a security review that flagged a lack of formal contracts and problematic use patterns [12][13]. Microsoft Copilot survived the cut, but only after tightened settings; some staff retain model access through Sweden's internal Intric platform [13][14].

This follows earlier blocks at the Swedish Tax Agency and SiS (the National Board of Institutional Care), with other agencies reportedly preparing similar moves [12]. The pattern is consistent: it's not a blanket rejection of AI, it's a rejection of AI without a procurement contract, data processing agreement, and audit trail behind it.

For Nordic public-sector watchers, this is the clearest signal yet that "just use the consumer app" is over in government contexts. The ones surviving the cut are the vendors who did the unglamorous work — contracts, compliance, data residency — before the headlines happened.