ZERADES Tech Daily — September 24, 2026

Agentic tooling stacked up today: Docker Cloud Sandboxes and the Sandbox Kit Spec, Cursor Rollouts + Security Review bots, JetBrains Air; open weights with MiMo-V2.6 and FLUX 3 Action.

24 September 2026

Summary: Agentic development infrastructure intensified today: Docker announced laptop–cloud microVM sandboxes and the OCI-based Sandbox Kit Spec (to CNCF); Cursor shipped deploy-health and PR security bots; JetBrains unified the IDE–team–governance layer under JetBrains Air. On the open-weights side, Xiaomi MiMo-V2.6 and Black Forest Labs’ FLUX 3 Action model stood out. The day was less about model leaderboards and more about answering “where do I run the agent, with what permissions, and how do I verify it?”

Highlights

1. Docker Cloud Sandboxes: start the agent on your laptop, finish in the cloud

Docker announced Cloud Sandboxes (24 September 2026), moving its previously announced microVM-based Docker Sandboxes onto Docker-managed compute. Same isolation model and CLI, with one-command local ↔ cloud moves such as sbx move … --to cloud; ready Kits for Claude Code, Codex, Copilot, Antigravity, Open Code, and Hermes; a shared MCP gateway; secrets injected via proxy (so the agent never sees raw keys); and network policies. Pricing is per-second compute (paused sandboxes are free); local sandboxes remain free.

Why it matters: Long-horizon (multi-hour) coding-agent jobs hit laptop sleep and battery limits. One isolation model plus portability is a practical bridge for teams looking for a “safe agent runtime” standard.

Source: https://www.docker.com/blog/introducing-cloud-sandboxes-start-on-your-laptop-finish-in-the-cloud/

2. Cursor: Rollouts + Security Review bots

In its 23 September 2026 changelog, Cursor announced two bots (Teams/Enterprise): Rollouts attaches to a PR and reports post-deploy health per environment (verified / regression / inconclusive); it writes a monitoring plan as a PR comment, checks Datadog and other telemetry, and on regression offers author alert / revert PR / handoff to a cloud agent (no autonomous merge/rollback). Security Review consolidates exploitable findings on every PR (injection, authz bypass, secrets, SSRF, unsafe deserialization, dependencies with known CVEs, etc.) in a single review comment; style/quality stays with Bugbot.

Why it matters: The “last mile” of agentic coding is no longer just generating code—it’s moving deploy verification and security review into a bot layer. It extends the shipping loop from the IDE to production signals.

Source: https://cursor.com/changelog/rollouts-and-security-reviewer

3. JetBrains Air: a product system for agentic development

On 22 September 2026 JetBrains announced JetBrains Air: from an individual workbench to a broader system where agentic work is started, run, coordinated, reviewed, and governed. Three surfaces: Air in JetBrains IDEs (agent steering and output verification), Air Teams (developer + autonomous agent workflows), Air Governance (formerly JetBrains Central; policy, visibility, audit, cost). Multi-vendor by design; Agent Client Protocol (ACP) and the ACP Registry connect IDE and agent harness. Junie will be supported across Air surfaces.

Why it matters: Not “another coding agent”—organizational governance and multi-vendor reality are at the center of the product strategy. The thesis that the IDE window isn’t the whole system overlaps the same week as Docker and Cursor announcements.

Source: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/

AI and models

Xiaomi MiMo-V2.6 open weights + RL stack

On 22 September 2026 Xiaomi open-sourced the MiMo-V2.6 series (Pro, Flash) and MiMo-V2.6-Distill-Qwen-9B; Hugging Face collection, technical report, 7k+ RL task environments, and an end-to-end RL training framework (environment interaction, trajectory, reward, policy optimization) were shared. The company claims Pro ranks strongly among open models on the Artificial Analysis Intelligence Index, with gaps still vs closed models (Claude Fable 5.1, GPT-6 Astra claims)—readers should verify scores independently. MIT-licensed weights on HF.

Sources: https://mimo.mi.com/docs/en-US/news/latest/v2-6 · https://huggingface.co/collections/XiaomiMiMo/mimo-v26

FLUX 3 Action: 7B world action model (open weights)

Black Forest Labs announced FLUX 3 Action on 23 September 2026: camera frame + text instruction → next ~2 s of action. 7B; DROID fine-tune claims 42.92% success in RoboLab (vs 36.8% for open 16B Cosmos 3 Nano policy in their table). LeRobot integration; SO-101 checkpoints; game/drone fine-tune examples. Weights under FLUX Kommunity License v1.0; code at github.com/black-forest-labs/flux-action.

Source: https://huggingface.co/blog/black-forest-labs/flux-3-action

Developer and software

Docker Sandbox Kit Spec → CNCF

The same day (24 September) Docker announced at WeAreDevelopers the Docker Sandbox Kit Spec under Apache 2.0, moving governance to CNCF. A Kit is an OCI image extension: agent + tools + a typed list of what it may access (host, credential, volume). A pinned image fixes the agent and permission requests together; registry/scanner/signing work with existing toolchain. Docker Sandboxes is the first enforcing runtime; emphasis on avoiding a single-vendor standard. Repo: docker/sandbox-kit-spec.

Source: https://www.docker.com/blog/docker-sandbox-kit-spec-cncf/ · https://github.com/docker/sandbox-kit-spec

Cursor bots (detail)

Rollouts: Origin/GitHub, CD deploy events, Datadog and other telemetry; feature-flag integration “coming soon”. Security Review: skips draft PRs; team rules for project-specific policies. First 10 days: Teams ~50 / Enterprise ~500 change trial credits for Rollouts.

Source: https://cursor.com/changelog/rollouts-and-security-reviewer

JetBrains Air / ACP

ACP aims to standardize the harness between IDE and agent for planning, reasoning, tools, model routing, and observability. Air Governance’s cross-provider cost and visibility claims map directly to enterprise multi-agent stacks.

Source: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/

Research and open source

MiMo-V2.6 RL suite: 7k+ environments (SWE, vulnerability reproduction, knowledge-intensive, web design), Distill-Qwen-9B starting point, verl / uni-agent / mini-swe-agent-based framework—a reproducible stack for agentic RL research. https://huggingface.co/collections/XiaomiMiMo/mimo-v26
FLUX 3 Action + LeRobot: Public robot and game/drone fine-tune recipes; expands open-weight options on the embodied/world-action path. https://huggingface.co/blog/black-forest-labs/flux-3-action · https://github.com/black-forest-labs/flux-action
Sandbox Kit Spec: Agent permissions as a portable artifact—the “authority portable” counterpart to container images. https://github.com/docker/sandbox-kit-spec

Industry

Major company news in this window skewed toward developer infrastructure (Docker, Cursor, JetBrains) and open models (Xiaomi, BFL). No separate “M&A / regulation” headline; short note instead: AWS Elastic Beanstalk Cluster Mode (shared EKS infrastructure) official What’s New entry 17 September 2026—InfoQ’s 23 September summary is secondary coverage; primary source: https://aws.amazon.com/about-aws/whats-new/2026/09/elastic-beanstalk-cluster-mode/

Short notes

Ansible Dev Tools v26.9.0 — 23 September 2026 forum announcement; ansible-dev-tools 26.9.0 on PyPI. https://forum.ansible.com/t/release-announcement-ansible-dev-tools-v26-9-0/46349 · https://pypi.org/project/ansible-dev-tools/26.9.0/
Semaphore sem-ai 0.4.0 — more compact pipeline responses (~99% smaller claim in demo), context switching, pre-flight checks API; Claude Code/Codex plugins. Update notes 21 September. https://semaphore.io/blog/sem-ai-smaller-responses-safer-access-and-pre-flight-checks
Docker Kit ecosystem partners (blog): Kits for AWS, Box, Datadog, Dynatrace, JFrog, Snyk, etc.—we have not yet verified independent runtime diversity; spec + first enforcing runtime from Docker. https://www.docker.com/blog/docker-sandbox-kit-spec-cncf/
Cursor Rollouts trial credits — limited free trial for Teams/Enterprise for 10 days after announcement. https://cursor.com/changelog/rollouts-and-security-reviewer
MiMo Desktop + UltraSpeed — desktop client and Pro UltraSpeed with V2.6 (blog claims “20x inference”; no independent measurement). https://mimo.mi.com/docs/en-US/news/latest/v2-6

ZERADES take

A clear theme in this first edition: a shift from the model race to the runtime and verification race. Docker (isolation + permission artifacts), Cursor (deploy health + security review), and JetBrains (IDE + team + governance) answer the same question at different layers: the agent can generate—but how does the team run it safely, audibly, and sustainably? On the open-source side, Xiaomi’s RL environment/framework sharing and BFL’s world-action weights support moving from “checkpoint only” to “train–deploy–measure” kits. Practical takeaway for ZERADES Labs: when choosing an agentic coding stack, evaluate sandbox, secret proxy, PR security bots, and cost/governance in the same sprint as the model.