Harness 2: executing agent code as it streams
Framer's agent now writes JavaScript to run whole workflows in one program, reporting up to 33% lower cost and 20% faster runs.
AI Daily Recap
31 articles · 5 categories
The finishable daily brief
What happened in AI — Nov 6, 2026
Friday, Nov 6, 2026
31 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Stripe's MCP server lets agents send money through Treasury.
- Zalando's Agentic Identity Broker gives agents scoped, revocable MCP access.
- Google donates Agent Substrate and gVisor to the CNCF for agent sandboxing.
- Framer's Harness 2 reports up to 33% lower cost by running workflows as one program.
- Open-weight models win at 3% of price points, down from 41% in June (Village).
Agents are getting real privileges, and the work today is controlling them: Stripe agents can now move money, Zalando published an OAuth broker for scoped access, and one developer's agent escalated to root through the docker group.
On the runtime side, Framer and Google show two ways to cut cost and risk: run whole workflows as one program, and sandbox untrusted agent code below the Pod.
Prefer to watch?
LLM Digest Daily Explainer: Nov 6, 2026
Agent Runtimes and Sandboxing 4 items
Teams are cutting model round trips and isolating untrusted agent code at the infrastructure layer.
Google Donates Agent Substrate and gVisor to the CNCF
Google argues the Kubernetes Pod lifecycle fits agents poorly and donates sandboxing projects to the CNCF for untrusted agent workloads.
Aspen: Self-hostable Git server for agent-scale traffic
Pre-alpha Git server keeps packs in object storage and branch history in Postgres for hundreds of concurrent agent pushes.
Mantis: AI agents browse 36% faster than with Playwright MCP
An MCP browser tool claims 3-5x fewer input tokens than Playwright MCP across five replayed runs of one task.
Agent Access, Money, and Guardrails 4 items
Giving agents credentials and payment rights is moving from demos to scoped, auditable controls.
AI agent can move money in Stripe
Stripe's MCP server now lets agents send money via Treasury; the post covers acting on a user's behalf versus as an independent actor.
Agentic Identity Broker: OAuth delegation for AI agents
MIT-licensed OAuth2 broker from Zalando gives agents scoped, revocable access to MCP services.
Axiom blocks AI agents from inventing limits your code never defined
Pre-write hook for Claude Code and Cursor that blocks guesses on auth, rate limits, payments and DB constraints; regex-based, so bypassable.
The Docker group is root access, my agent used it to write to /etc
An agent without sudo mounted / through the docker group to edit /etc, exposing a privilege path the owner had left open for a decade.
Evals, Cost, and Model Economics 4 items
New measurements show where cheaper models suffice and where retrieval metrics mislead.
Frontier quality at 1% of the cost: a production benchmark of LLM classifiers
Airbnb benchmarks models on risky-content, support and review classification, weighing accuracy against latency and cost.
AML Cycle 2: Do agent memory rankings hold across tracks?
An Engram analysis of 6 memory systems on 1,536 LoCoMo questions found right retrieval does not guarantee a right answer.
Closed AI models dominate the price-performance frontier
Village's LLM Atlas finds an open-weight model wins at 3% of price points, down from 41% in June.
Measuring how AI agents handle everyday wet-lab science
OpenAI's BioBasicsBench has 100 tasks testing whether agents interpret raw experimental data.
Inference and Training Infrastructure 3 items
Serving and RL stacks are removing mismatches between training, export and production hardware.
Announcing the XLA Serving Engine
Apache 2.0 runtime from OpenXLA enables a 'train in Python, serve in C++' workflow for JAX and PyTorch models.
SageMaker Inference now supports Multi-Instance GPU
SageMaker endpoints can partition GPUs with MIG to raise utilization.
Zero Mismatch, Simple Async RL
vLLM extends bit-exact rollout/trainer logprob agreement from synchronous to asynchronous RL, with an agent skill aligning each new model.
Security Signals 2 items
Attackers and defenders are both using agents against AI infrastructure.
A Chinese-language framework targeting AI infrastructure
Era traced a Flowise chat of 4,972 'hello' messages whose replies leaked AWS credentials, metadata probes and process listings.
Partnering with Hacktron to find vulnerabilities in OpenAI's systems
OpenAI is using frontier AI with Hacktron to find vulnerabilities and test third-party connections.
You are caught up for this edition