🚨 NVIDIA's agent-in-one-Python-class, a 2.6B agent running on your phone, and a harness that beat the human baseline on ARC-AGI-3

Presented by Mend.io  Securing AI agents, MCP servers & LLM apps — the practical framework for finding, triaging, and protecting agentic AI in production. Get the guide →

Every story this week hands the operator something that used to live with the vendor or inside the weights. The deployment boundary. The moderation policy. The team’s memory. The undo button. Even the harness is now state you can edit. Hold the model constant across these releases and the behavior still changes — the differentiation this week lives in the substrate, not the weights.

93.3%

RULER at 10M tokens

92.1%

test tasks vs 78.9%

5×

faster fork than Docker

01 · Pokee AI · Models

Pokee AI Releases Pokee-Isaac 28B: A 10M-Token Context Agentic Model Built to Run Inside the Customer Boundary

Long context and long-horizon coherence have been cloud-only capabilities — which locks out anyone whose data cannot leave the building. Pokee-Isaac is a 28B text model with a 10M-token window built to run where that data lives: VPC, on-premises, or on-device.

On Pokee’s internal panel it scores 93.3% on RULER at 10M tokens, where every baseline returns 0.0 beyond 2M. Prefill reaches 137,200 tokens/s at full context on one B200 with decode flat near 335 tokens/s, and it leads BFCL v4 at 70.94. Weights are not published — deployment is licensed in, with provisional list pricing at $0.15/$1.00 per million tokens.

Read on Marktechpost →  · Technical report

02 · Northeastern + Stanford · Agent runtime

Meet Shepherd: An Open-Source Python Substrate That Lets Meta-Agents Fork, Replay, and Revert Any Agent Run

When an agent misreads a traceback at step 10, your recovery options are pay to patch forward or pay to restart from step one. Git versions files, not a live process or a warm cache. Shepherd records the run itself as a Git-like trace of typed events — so “go back to step eight” is a single fork.

Each interaction is effectively a commit covering the agent process and the filesystem together, copy-on-write. The research team reports forking 5× faster than Docker and over 95% prompt-cache reuse on replay, with core operations mechanized in Lean. A supervisor meta-agent built on it lifts pair-coding pass rate from 28.8% to 54.7% on CooperBench.

Read on Marktechpost →  · Paper  · GitHub  · Project

03 · Mistral AI · Safety

Mistral AI Releases Shieldstral 1.0 3B: An Open-Weights Policy-Adaptive Multimodal Safety Classifier Matching Models 7× Its Size

Most guardrail models bake a fixed taxonomy of harm into the weights, so re-targeting one means retraining. Shieldstral inverts that: you write the policy as a plain-language question at inference time and get a calibrated safety score from a single forward pass.

Built on Ministral-3-3B-Base-2512 with a native Pixtral vision encoder, Apache 2.0. It reports 84.9% average F1 on text safety — matching GPT-OSS-Safeguard-20B — and 83.8% on multimodal. Honest weak spots: policy adaptability trails the 20B (91.3% vs 94.1%), and multilingual prompt classification lags on Arabic and Indonesian.

Read on Marktechpost →  · Announcement  · Hugging Face

Sponsored · Mend.io

Your agents don’t do what their code says. Secure them anyway.

Agent behavior isn’t defined by code alone — it emerges from models, prompts, retrieved context, and the tools an agent can call. Mend.io’s practitioner guide, Securing AI agents, MCP servers & LLM apps: A practical framework, gives your team the playbook to discover, triage, and protect agentic AI in production.

Inside: a five-layer agentic attack surface map, a 12-point misconfiguration checklist, evidence-based triage, and runtime guardrails that ship as a Python SDK or a standalone Docker API server — aligned to NIST AI RMF, OWASP, ISO/IEC 42001, and the EU AI Act.

Get the guide → · Our breakdown on Marktechpost

04 · NVIDIA · Frameworks

NVIDIA AI Releases NOOA: An Object-Oriented Python Framework That Turns an AI Agent Into a Single Python Class

Agent development today is split across prompt templates, tool schemas, callbacks, and workflow graphs. NOOA collapses all of it into one class: methods are actions, fields are state, docstrings are prompts, type annotations are enforced contracts. A method whose body is ... is completed at runtime by an LLM-driven loop.

NVIDIA reports 82.2% on SWE-bench Verified and 86.8% on CyberGym L1 at roughly half the tokens of the open harnesses it was compared against. Apache 2.0, pip install nooa (v0.0.8), Python 3.12–3.13 — and explicitly a research preview, with containers as the containment boundary.

Read on Marktechpost →  · Paper  · GitHub

05 · Microsoft · Coding agents

Microsoft Open Sources code-testing-generator: a Polyglot Unit-Test Agent That Hits 92.1% Task Completion Versus 78.9% for Stock Copilot

“Generate unit tests” doesn’t say which framework, file location, or assertions to use. This agent settles those decisions by reading the repository before writing anything — then plans, writes, runs, and checks its own tests before claiming done.

On Microsoft’s internal 152-task benchmark: 140 completed (92.1%) versus 120 (78.9%) for stock Copilot on the same model and prompts. The gain concentrates on vague prompts (88.8% vs 66.3%) and diff-targeted tasks (15/15 vs 0/15) — at effectively identical line coverage, 72.4% vs 72.2%. Ships in the MIT-licensed dotnet/skills repo, and it never modifies production code.

Read on Marktechpost →  · GitHub

06 · Tencent Cloud · Agent memory

Tencent Cloud Open-Sources TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents

If project context was already explained once, a new session shouldn’t need it repeated. v2.0 turns conversations, docs, and code into four governed memory assets — Chat Memory, Skill, LLM-Wiki, Code-Graph — that are versioned, permissioned, and equipped to specific agents. The new part is the governance: your teammate’s agent can read what yours learned, without leaking what you marked private.

Per the README, private is owner-only — not readable even by team admins — and new assets default to it. The proxy speaks both Anthropic and OpenAI protocols. MIT-licensed and self-hosted: three Docker images, one command, amd64 and arm64. Stable 2.0.0 landed August 3.

Read on Marktechpost →  · GitHub

07 · Liquid AI · On-device

Liquid AI Releases LFM2.5-2.6B: An On-Device Agentic Model With 128K Context, Tool Calling, And Open Weights

A background agent that runs continuously has one fatal flaw in the cloud: per-token cost. LFM2.5-2.6B is Liquid’s answer — a 2.69B-parameter agentic model that runs entirely on-device, in under 2.5 GB of memory, at 220 tokens/s on an M5 Max and about 30 on a phone.

30 layers — 22 double-gated short convolution blocks plus 8 GQA — 128K context, pre-trained on roughly 34T tokens. It beats gemma-4-E4B-it and Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct, while trailing Qwen3.5-9B on LiveCodeBench v6 (59.41 vs 69.86) — Liquid explicitly does not recommend it for agentic coding. Open weights under lfm1.0, with GGUF, MLX, and ONNX from day one.

Read on Marktechpost →  · Announcement  · Docs

Sponsored · Datalab

Marker v2: documents in, Markdown out.

PDFs, scans, DOCX, and PPTX → Markdown, JSON, and HTML — with tables, inline math, and reading order surviving the trip. Built for the pipelines feeding your RAG stack and your training data.

76.0 on olmOCR-bench at 2.9 pages/second of GPU throughput. Fully open source under Apache 2.0.

Try Marker v2 → · Our benchmark breakdown

One more thing…

Prime Intellect · Agent harness

Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel

Every harness this year has added tools. Prime Agent removes them. Models get exactly one: a persistent IPython kernel. Skills, tools, and sub-agents are pre-imported modules inside it — rlm("sub-task") launches a child session with its own model, kernel, and history. Long sessions stop being a context problem and become a programming problem.

The Continual Harness makes prompts, sub-agents, skills, and memory state the agent can create, read, update, and delete from its own trajectory — the base prompt stays immutable and any bad update reverts by ID. With Opus 5 it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%. MIT-licensed, one-command install on Linux or macOS.

Read on Marktechpost → · Announcement

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. Placements are written in your voice and placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for Mend.io →