What happened
The protocol for wiring models to tools went stateless
The 2026-07-28 Model Context Protocol specification dropped the requirement for persistent session state: an MCP server can now be a lightweight Python or TypeScript process answering over plain HTTP. A new protocol roadmap was published on top of that, and both AWS AgentCore Gateway and Cloudflare already support it.
Why it matters: until now, connecting a model to your own tools meant running a service with session memory, with everything that brings — sticky scaling, reconnections, state that goes bad. Without state, an MCP server deploys like any other endpoint. That puts it within reach of a small business that already runs an API, not just a company with an infrastructure team.
Specification 2026-07-28 · New roadmap
Docker shipped sandboxes to isolate the code the AI writes
Docker Sandboxes 0.39.0 landed on 19 August, built specifically to run agent-generated code in isolation. Alongside it came security patches for Docker Desktop and Engine 29, including CVE-2026-17106.
Why it matters: the moment an agent not only writes code but also runs it, isolation stops being good practice and becomes a requirement. That the tool comes from Docker itself rather than a side project is what makes it defensible in front of a client.
LangGraph unified how an agent gets built
Version 1.2.x settled on create_agent as the standard approach and advises against the classic AgentExecutor. In exchange it offers fine-grained state control, interruptions with a person in the loop, and deterministic graphs.
Why it matters: human-in-the-loop stops being a patch each of us builds by hand and becomes part of the framework. It is exactly the pattern I use for this radar — the agent drafts, I approve — and it reassures me to see it turned into a standard rather than everyone''s own craftwork.
Gemini 3.7 Flash went generally available, and older models were switched off
On 13 August gemini-3.7-flash reached general availability, tuned for software engineering, code execution and low-latency tool calls. Ten days later the definitive shutdown of Claude 3 Haiku on cloud platforms was confirmed, with migration mandatory.
Why it matters: less for the new model than for the deprecation. If you have automation in production leaning on a model that gets switched off, the work is not picking the best model: it is having written the code so that swapping it is one line of configuration. This is the second deprecation of the year.
Gemini API changelog · Deprecations
Further down the stack: Python starts running without the GIL
PyTorch 2.14 RC1 arrived on 14 August, advancing compatibility with Python 3.15 and adding experimental free-threaded support — no GIL — on Linux, for multithreaded training and inference. Scikit-Learn published 1.9.0 with faster matrix transformations. FastAPI settled app.frontend() for serving a single-page app without wrestling with static files. And Supabase added AND filters and column selection to Postgres Changes.
Why it matters: free-threading is the deep change here. Twenty years of "real parallelism in Python is painful" are starting to shift. It is still experimental, but once it lands as stable it changes how data pipelines get designed, not just how fast they run.
PyTorch · Scikit-Learn · FastAPI · Supabase
What changed since the last period
A week ago MCP was a freshly published specification. Today it has its own roadmap and two cloud gateways supporting it in production. That is the real jump of the period: people stopped arguing about what the protocol should look like and started building on top of it.
The rest moves in the same direction. Sandboxes to run what the agent writes, a framework that standardises where a person steps in, a model tuned for calling tools. Four pieces of the same puzzle, published by four different players in eight days.
What I am watching
- Whether Python''s free-threading reaches stable. It is the only item in this edition that changes how code gets written, not just which tool it uses. The rest is plumbing; this is the material.
- Whether sandboxes become the default rather than the tidy option. As long as running an agent''s code unisolated stays the easiest path, it will keep happening.
- The pace of deprecations. Two model shutdowns in a year sets a rhythm. Anything I build for a client has to be able to change provider without being rewritten.
Worth keeping in mind
A release is not an improvement until someone puts it in production and keeps it there. Half of what appears here will come to nothing, and it is not always possible to tell which half. This edition is a record of what moved, not a recommendation of what to adopt.