In partnership with
CopilotKit open-sourced an AI coworker you can self-host.
AWS let agents write a local copy of 200+ services.
Google built a model that only answers multiple choice.
▲ AI DEV SIGNALS
by Marktechpost · November 12, 2026
Hi there,
Most AI coworker products ask you to rent everything: the model, the hosting and the rules. This week, an MIT-licensed project went the other way. AWS let agents build a test double of its own cloud. Google shipped a model that never writes a word. And MBZUAI opened up a 29B world model for robots.
01 · Featured: the coworker you actually own
CopilotKit releases OpenBot 1.0: an MIT-licensed, self-hostable AI coworker with built-in governance
OpenBot is an agent workspace from the team behind the AG-UI protocol. You give a Bot a task, watch it work, and approve the next steps in the same thread. You bring the models, agents and infrastructure. Every action is checked against policy before it runs.
How it works: Each Bot works inside a container with its own Chromium browser, logins and files. It can only reach the tools you grant. Any AG-UI agent can join by pasting its URL, including LangChain, Mastra, Google ADK, Pydantic AI, LlamaIndex and AWS Strands. Bots always ask before deleting, sending, sharing or posting. Payments always stay with you, even inside routines.
The numbers: A Bot can split work across up to 5 subagents in parallel. Generative UI runs on A2UI 1.0 with 65 components. Hundreds of apps connect through Composio and Arcade. The repo just passed 10K GitHub stars. Desktop apps cover Mac (Apple silicon and Intel) and Windows.
The fine print: The desktop app needs Docker Desktop. By default, all of a person's Bots share 1 computer. Giving a single Bot its own isolated computer is a Business feature. SSO, named "boundaries" and the CSV audit trail are also part of OpenBot for Business, which is still in early access. Read the full breakdown.
Analytics on Live Data Without Leaving Postgres
When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.
TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.
CERN runs Postgres this way for sensor data from the Large Hadron Collider.
No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.
Start on Tiger Cloud and get $1000 in credits.
02 · 4 more drops: a cloud written by agents, a model that only scores
It's free and CC0-licensed, and the AWS CLI and SDKs talk to it instead of AWS. Autonomous agents wrote all of the code; humans built the tests and merge gates. It's built from 438 AWS API models, with 1 Rust crate per service. The catch: it passes 56.6% of 42,857 published fidelity tests, and CloudFormation and IoT lag behind.
It doesn't generate text. It scores up to 1,000 typed questions against 1 shared input and returns calibrated probabilities in a single forward pass. Output tokens are always 0. The target is the small branching calls inside agents: routing, policy checks, tool selection and grading.
You give it a scene and a text action, and it simulates the outcome as video, up to 20 seconds at 24 fps. Weights, training code and inference code are Apache 2.0. It leads 6 tested models with 34.0% full-plan completion on 100 multi-step tasks. The catch: that score is self-reported, and it trails MiniMax-H3 on single actions and human preference.
IBM plans to link 2 r3 chips into a 1,040-qubit system in 2027. The r2 chip available today runs 100,000+ circuits per second, about 25x Heron. The catch: r3 is an announcement; r2 is what you can run on IBM Quantum Platform now.
03 · Signals: a 1.6T download, a library that deleted itself, and agents that ignore robots.txt
LongCat-2.5 weights are now on Hugging Face. Meituan's 1.6T MoE (about 48B active) is MIT-licensed. Its self-reported 88.4 on Terminal Bench v2.1 compares with 89.1 for Claude Opus 5.
Browser Use 1.0 deleted 80% of its own code. It removes 136,866 lines; the agent now writes Python directly on Chrome DevTools Protocol. It scores 62.5 vs 56.3 on BU Bench v2, at $0.070 vs $0.082 per task.
n8n ships Nod, a personal agent with its own sandbox. It's an Apache 2.0 research preview for Mac and Windows from n8n Labs. Use your own model provider or a local model.
Most agents ignore no-scraping rules. Across 1,626 sites that prohibit scraping, ScrapeBench finds most agents comply less than 5% of the time. On sites that allow it, Claude Opus 4.6 refused 98.5% of the time.
Low attack rates can hide heavy over-refusal. ErosionBench's 1,728 scenarios show perimeter defenses refusing 54–92% of legitimate work. Its reference auditor reaches 98.5% recall at a 2.9% false-positive rate.
MCP agents keep going when the goal is impossible. On MCPHallu's Unreachable Goal tasks, only 23.2% of runs pass. Code is on GitHub.
1 question
If your agent had to ask before every send, share and delete, how many approvals would you click through each day?
Have a great Thursday.
until next week,
the marktechpost team
REACH 70K+ AI BUILDERS
Was this email forwarded to you? Sign up here