Haiku 5.5 cuts small-model cost ~75%, Windows ships agent sandboxes, and debugging eats the week

Strands Box — is AWS's open-source sandbox for agents: it locks the agent into an operating-system box, then checks every file read, shell command and outbound web call against rules you write. It is an off-the-shelf answer when a client asks how an agent can run unattended without being handed the whole machine. Rules can count history — let the agent post incident updates to Slack, but no more than three times in ten minutes. It is a developer preview, macOS first, Apache 2.0. AWS Open Source Blog

Docker Agent — is a Docker command-line plugin that describes an agent in a YAML file instead of code, then ships it through a registry like any container image. Clients already standardised on Docker get versioned, reviewable agent definitions without adopting yet another platform. Pre-installed from Docker Desktop 4.63 onwards, works with OpenAI, Anthropic, Gemini, Bedrock and local models, and the repo sits at roughly 3.6k stars. GitHub

Mecatl — is an open-source agent harness designed to run in the cloud rather than on a laptop, from Stacklok, the company founded by two Kubernetes co-creators. It is the first credible answer for enterprises that want agent sessions governed centrally, the way they govern email, instead of scattered across engineers' desktops. The agent loop is kept separate from tool execution, memory and model provider; the paid piece is the control plane that ties identity, policy and audit together. Latent Space

“We think, in the future, interactions that you have with an agent while you’re working for an enterprise belong to that enterprise”

OpenHands Agent Canvas — added Agent Profiles, which pin down exactly which tools, connectors and secrets a given agent may touch, and gave every conversation its own isolated Docker container. If you run triage, coding and review agents over one repo, this is the scoping-and-isolation story a client's security team will ask you for. Eight releases in September, v1.17.0 through v1.24.0, with 49 outside contributors and nearly 600 merged pull requests. OpenHands

Intelligent UI — is OpenAI's new answer format: rather than paragraphs, ChatGPT builds tappable buttons, little calculators, maps and editable charts inside the reply itself. Your clients will see this within days and start asking why the agent you built for them still returns a block of text. It rolled out Wednesday with a new GPT-6 model for Pro, Plus, Business and Enterprise users, reaching free and Go tiers a day later, and users can dial the visuals back. TechCrunch

Microsoft Execution Containers — Windows 11 now ships an operating-system feature for sandboxing AI agents, announced alongside Nvidia RTX Spark laptops that run large models on the device itself. A standard-issue corporate laptop becoming a governed place to run agents changes the 'where does this actually execute' conversation more than any model release this week. Surface Laptop Ultra starts at $2,600, the Surface RTX Spark Dev Box at $6,000, and Microsoft will knock up to $1,000 off for a traded-in MacBook Pro. NVIDIA Blog

Nous Research — raised $90 million at a $1.5 billion valuation and is taking its open-source Hermes agent into companies as a product called Hermes for Businesses. An open-source agent with genuine volume behind it is now something clients can name in the same breath as the labs, especially where data has to stay in-house. Hermes has been cloned more than 24 million times and the company estimates it drives about 2.5% of global AI token usage; total funding reaches $158 million. TechCrunch

Personal Agent Protocol — Meta, Walmart and Bret Taylor's Sierra are writing shared standards for how a shopper's personal agent identifies itself to a business, connects to an account and is told what it may do. Clients wondering what to expose to agents arriving on their site now have a candidate standard to point at — and a good reason not to wait for it before deciding their own rules. Analysts note it builds on OAuth, the existing permission plumbing, but that neither OpenAI nor Anthropic has signed on, which is what would make it real. Techstrong.ai

Reported: Meta and Microsoft push staff off Claude. The Information reports, and this is not confirmed by either company, that both are steering engineers towards their own coding tools. Microsoft's forecast internal spend on Anthropic has reportedly dropped by over a third, with per-employee monthly ceilings cut from $100,000 to around $10,000 in its cloud and AI group. Even the most AI-committed buyers are starting to treat agent spend as a budget line with a cap, which is the posture your clients will copy. Meta's Claude Code users are said to have roughly halved to about 30,000, yet the company still reportedly spent over $105 million on Claude Code in one 28-day period. RS Web Solutions

AI Workflow Factory — packages ServiceNow's process mining, app engine and control tower into one path from finding a process worth automating to running and auditing the agent that does it. This is the shape of the enterprise pitch you will be bidding alongside or against — Infosys and Accenture are already folding it into their practices. The companion product, Autonomous Engineer, writes and tests implementation work after a human approves the plan, and is currently invite-only early access. Techstrong.ai

Agentic Value Model — is AWS's argument that the old hours-saved sum, designed for rule-based automation, misses most of what agents are worth, replaced by four value pools each with a named owner and a realisation test. It is a ready-made structure for the ROI slide clients keep asking you to produce, including the uncomfortable line that freed hours only count if headcount or spend actually falls. It cites McKinsey: a dollar on the technology needs three on process redesign and five on adoption, and under 10% of enterprises have scaled agents to tangible value. AWS Machine Learning Blog