Realtime AI News
Open-source Agnes Harness splits agent decisions between a 9B model and an LLM
QbitAI detailed Agnes Harness, an Apache 2.0 open-source agent system with 2.1k GitHub stars that routes routine next-step decisions to a 9B model called Agnes Sys1.0. The vendor claims judgments about 14.3 times faster and roughly 1/27 the per-call cost of a large model, with test-time training and recursive self-improvement to retain execution experience.
On November 20, Chinese AI outlet QbitAI published a detailed look at Agnes Harness, an open-source agent system released under the Apache 2.0 license that allows commercial use and has gathered 2.1k GitHub stars. The project rests on two mechanisms: a dual-system decision layer called Sys1, and a two-layer self-evolution loop combining test-time training with recursive self-improvement. The team also published technical reports for Agnes Sys1.0 and for the Harness itself.
The system targets a cost problem that has become familiar in enterprise agent deployments. QbitAI opens with an estimate from Zaniel, an independent forward-deployed engineer and Stanford PhD student: software used to cost 90 units to build once and only 10 to adapt for each customer, but AI pushed development down to about 1 unit while that adaptation bill barely moved. Adaptation is expensive because an agent has to relearn each customer's business processes, tool stack and on-site equipment.
Sys1 splits the decision workload. Complex reasoning stays with a large language model acting as System 2, while high-frequency, well-defined next steps go to a purpose-trained 9B model, Agnes Sys1.0, acting as System 1. It plugs into the Harness and decides which tool to call, whether to wait or act immediately, and when to escalate to the large model.
The vendor's comparisons are blunt. In a simulated home-flood task, the large model needed 7.6 seconds on average to decide the next step versus 0.53 seconds for Agnes Sys1.0, roughly 14.3 times faster. In a luggage-room test covering weighing, vertical placement and temporary storage, 100 decisions cost about $0.00247 per call on an online large model versus $0.000091 for Sys1.0, about 1/27. In a robot-dog simulation where the agent had to find a door and pass through it, pairing the large model with Sys1.0 cut total task time by 32.9%.
The second mechanism addresses learning. TTT lets Agnes Sys1.0 inspect a new task, generate its own practice questions with known answers, and then tune only a small set of end-layer calibration parameters rather than the 9B backbone. The twist is deliberate amnesia: those temporary parameters reset when the task ends, so experience from one device or job does not bleed into the next.
RSI is what keeps the experience that matters. After each run, the Harness replays the episode, looks for repeated errors and time sinks, and adjusts how prompts, memory, skills and subagents work together; candidate fixes are validated on the current task before being versioned into a playbook for reuse. In one equipment check-in task, the agent failed its first attempt by reading a scale before it stabilized and timed out after 30 minutes, then followed the validated procedure on the second run and finished in under eight minutes.
Reaching the physical world relies on plugins and a hardware standard. Business tools, model adapters, agent workflows, skills and even interfaces can be packaged as plugins, and existing systems connect through MCP. MHS, the Model Hardware Standard, has devices describe their state, available actions and returnable data in a common format so the Harness can sequence them. A QbitAI writer deployed it locally on a MacBook Air with Node.js 24.10+ and pnpm 10.34.5 in about 50 minutes and used a self-built review plugin to flag six problems in a deliberately flawed draft. The project states that real-time control and physical safety remain with device controllers.
On capability, the team reports a nine-benchmark self-assessment in which the Harness placed first against its comparison group on six and tied Codex on one, scoring 87.07 on DeepPlanning and a 40.69% pass rate on BigCodeBench-Hard V1 — vendor numbers that still need independent replication. For the agent industry, the more important question is whether replication cost can be amortized: if validated workflows and documented failures become the starting point for the next deployment, the last, stubborn slice of adaptation cost finally shrinks. Watch the MHS plugin rollout in the next release and whether outside teams reproduce these results on real production tasks.
Why it matters
If outside teams reproduce these figures on real production tasks, dual-system routing plus versioned playbooks could turn agent deployment from one-off engineering projects into reusable delivery capacity.
Nearby Updates
All11/20, 10:55
Mechrevo launches Master Road X10 enterprise desktop AI host
Mechrevo has announced the Master Road X10, an enterprise desktop AI host, according to a China.com report. The headline-level information covers the name and business positioning only; specifications, pricing and availability have not been disclosed.
11/20, 10:36
Kimi-K2.5 lands in Moonshot AI’s official Hugging Face repo as an image-text model
Moonshot AI has refreshed its official Hugging Face repository with Kimi-K2.5, a model whose pipeline is tagged image-text-to-text and whose weights ship in safetensors format. The repo reports 391,280 downloads and 2,902 likes, giving the new version a public multimodal footprint before any technical report appears.
11/20, 10:34
Hainan releases its first filed tourism LLM as Yuduoduo agent lands in Rongshanliao
Hainan released its first filed large model for the culture and tourism sector, with an agent called Yuduoduo deployed in Rongshanliao on November 18, according to Guandian. The launch puts a locally filed model and agent into a concrete travel scenario rather than a generic chatbot.
11/20, 09:25
Korean exam test: GPT and Gemini score near-perfect, domestic K-EXAONE slips on inquiry section
On South Korea's national college entrance exam, GPT and Gemini turned in near-perfect answers while the country's top domestic model, K-EXAONE, lost ground on the inquiry section. The comparison puts the gap between frontier models and a sovereign Korean model onto a single, publicly legible scorecard.