$ today_in_10_seconds
Alibaba shipped a 2.4-trillion-parameter model whose headline demo isn't a benchmark — it's a repo it maintained by itself for sixteen days. Meanwhile the EU AI Act's enforcement powers switched on Sunday, and the obligations most vendors warned you about aren't the ones that landed. Capability got longer-horizon this week; everything else that shipped was plumbing to watch it.
$ lead_story
Qwen3.8-Max claims a 16-day autonomous coding run
Alibaba released Qwen3.8-Max on Monday — a mixture-of-experts model with 2.4 trillion total parameters, built on the Qwen 3.5 architecture, with a context window up to 1 million tokens — and confirmed open weights ship next week alongside a smaller Qwen3.8-27B checkpoint. The pitch is long-horizon autonomy, not single-turn cleverness: in the company's own test, the model spent roughly 16 days building and maintaining a project called oh-my-cli, producing 265 commits, 127 pull requests and 151 issues through a loop of task intake, implementation, testing and repair, with every merge gated by end-to-end tests and CI checks. On price, Alibaba opened global API access at roughly 40% of Claude Opus 5 for input tokens and 24% for output — which is the part that actually changes your architecture, because it makes leaving an agent running overnight a rounding error instead of a budget line. The skeptics landed on the right objection: a 16-day agent run changes the unit of risk, because a small mistake compounds across thousands of actions, and enterprise value depends on whether teams can inspect checkpoints, trace decisions, cap permissions and roll work back before an error reaches production. Note also what's deployable and what isn't — the hosted API is OpenAI- and DashScope-compatible, so integration is a base-URL swap, but a 2.4T checkpoint is a multi-node datacenter artifact. The 27B is the one that fits your GPUs.
// why_it_matters: When an agent run outlives your sprint, your rollback story becomes your product.
$ also_worth_attention
The EU AI Act deadline that actually landed on August 2
Article 50's transparency and information obligations became generally applicable and enforceable by national competent authorities across the EU on 2 August 2026, covering any product that talks to users or generates images, audio, video or text, regardless of high-risk classification. The obligations everyone spent 2025 preparing for did not arrive: via the Digital Omnibus on AI, stand-alone Annex III systems now have until 2 December 2027, and AI embedded in Annex I regulated products until 2 August 2028. What did switch on alongside Article 50: enforcement powers over general-purpose AI and the full penalty regime. The practitioner read — if a vendor is still selling you August 2 as your high-risk cliff, they haven't updated their deck since June, and the deferral is runway, not relief.
link →CopilotKit open-sources Channels SDK under MIT
Channels drops an existing agent into a messaging platform without a per-platform rewrite, connecting over AG-UI — so LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK, or a plain HTTP agent you wrote yourself all work unchanged. Slack and Microsoft Teams are the early access surfaces, with Discord and Google Chat planned. Messages are described once and rendered as native Slack Block Kit or Teams Adaptive Cards, with buttons, choices and approval gates placed directly in the conversation before an agent acts. The architectural detail worth stealing: your agent, tools and business logic stay in your infrastructure while platform credentials stay out of your agent process. Approval gates rendered in the surface where people already work is a better human-in-the-loop pattern than a dashboard nobody opens.
link →Silicon Motion pitches KV cache offload as agent infrastructure at FMS 2026
At FMS 2026 in Santa Clara (August 4-6), Silicon Motion is showing storage aimed at what it calls the shift beyond LLM training toward agentic AI, where real-time inference depends on fast, predictable data access. The headline part is the SM8366 MonTitan SSD reference design kit, positioned explicitly for KV cache offload and agentic AI infrastructure, with capacities from 32TB and 64TB on TLC up to 256TB on QLC, and configurable over-provisioning to 28% supporting up to 3 DWPD. Its keynote also introduces PerformaShape, a controller technology pitched at intelligent workload isolation and predictable performance for multi-agent AI workloads. Yes, it's a trade-show press release — but read it next to the lead story. Sixteen-day agent runs generate enormous cache state, and noisy-neighbour contention between agents is a real production failure mode. Hardware vendors don't invent categories; they follow spend.
link →
$ bigger_picture
The trend behind the headlines
Look at what shipped this week versus what got funded. The capability news is one item — a cheaper model that runs longer. Everything else is control surface: transparency rules with teeth, approval gates in Slack, workload isolation in the SSD controller, and groundcover's $100M Series C for observability that keeps telemetry inside the customer's own cloud, from a company whose CEO frames the round as changing how engineers and agents collaborate to bring true autonomy into production operations. That's not four unrelated stories. It's the market pricing the same bet: model quality is converging and commoditising, while the ability to trace, cap, contain and roll back an autonomous run is where differentiation now lives. A 16-day agent that you can't checkpoint isn't an asset, it's an unbounded liability with good benchmarks. Build the brakes before you buy more engine.
oh-my-cli
The repo Qwen3.8-Max spent sixteen days building, now public. Skip the novelty and read it as a containment reference: it treats all discovered instruction files strictly as data that can never activate tools or override a safety boundary, assigns trust classes so an out-of-workspace instruction can never override the workspace's own policy, rejects symlinked instruction files whose real path escapes their directory, and keeps secrets and host paths redacted. Its autonomy contract, policy files, workflows and CODEOWNERS form a protected governance plane the bot may propose changes to but cannot apply. That last line is the design pattern most agent deployments are missing.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime