Yesterday’s dispatch was workhorse economics — Gemini 3.7 Flash half-price intro + Grok 4.6 tying Sol on the AA Index. Day before that was open-weight coding week. Today is product ops and harness plumbing: Microsoft’s consumer Copilot features actually cut over on August 18, and the multi-agent story worth wiring is not another flagship — it is AgentRadio, Coral’s passive-awareness bus that nearly doubles long-horizon codebase QnA for four Claude Code workers.

Lead: cutover day, not another teaser

Microsoft is merging the consumer Copilot surface with the Microsoft 365 Copilot app into one experience and retiring features that did not earn “the right to exist” (internal framing reported earlier this month). Primary sources: TechCrunch, Microsoft support — Updates to Copilot, GeekWire on the merger path.

Personal vs work/school accounts stay separate under org controls. Chat history and most content migrate; mobile needs the updated app. Files from the stand-alone consumer app move toward OneDrive during the transition. Microsoft’s pitch is simpler navigation and one place to jump from chat into Word/Excel/Outlook — the same consolidation wave as Claude folding Cowork into Chat and OpenAI folding Operator into ChatGPT.

Bold insight: big-lab AI product strategy in mid-2026 is subtractive. The bill for “ship every demo” showed up as confused surfaces. Cutover day is the real story — not another model card.

If you are retooling a home lab or a small team desk after the Copilot shuffle, the boring kit still wins: a solid USB-C dock so the agent laptop is not fighting power, noise-cancelling headphones for long agent review sessions, and a UPS so a brownout does not corrupt a mid-migration OneDrive sync.

Secondary: AgentRadio — listen while you work

The other thread operators should actually try this week is coordination latency, not parameter count.

Coral AI Labs (with SnT Luxembourg / King’s / Hull co-authors) shipped AgentRadio: an asynchronous message-passing layer so coding agents can keep working while peers drop discoveries mid-task. Paper: arXiv:2607.28430. Code: github.com/Coral-Protocol/AgentRadio (Apache 2.0). Field writeup: VentureBeat.

Three primitives, deliberately dumb:

1. create_thread — open a conversation among participants
2. send_message — non-blocking append
3. wait_for_mention — background watcher; deliver message + thread snapshot without freezing the main loop

Requirement: the harness can run a shell command in the background (Claude Code / Codex-class loops). No fork of the model weights — just a message server + thin adapter + prompt that keeps one watcher alive.

On SWE-Atlas QnA (long-horizon natural-language questions over live production repos — you must run the software, not only greps):

That is the spicy claim: four older Opus 4.6 workers with passive awareness beat one newer Opus 4.8 solo on this benchmark. The MinIO case study in the paper is the engineering parable — two agents independently realized they needed per-request server logs; without async share, one gave up quietly and the team locked the wrong answer at review. With AgentRadio, one broadcast hit the shared worklog while peers were still executing, and the score flipped to full marks.

Limits they own in public: passive awareness distributes ideas someone already formed — it does not invent the missing negative hypothesis (Grafana case). Faster gossip can also spread a shared wrong assumption. Next bottlenecks they name: attention governance, evidence-aware routing, provenance for “which agent claimed what.”

Commercial cousin called out in coverage: Coral Code — bottom-up specialists spun from the repo rather than a fixed four-agent protocol on every ticket (coralcode.dev).

Tertiary: contain the worker, not just the model

Same week ops stack: Hazmat — open-source containment that runs coding agents (Claude Code, Codex, OpenCode, Cursor Agent, custom scripts) under a separate OS account with project-scoped write and credentials outside reach. Pre-launch printout of write/read/network terms; macOS sandbox + firewall path; demo that proves a key under $HOME stays unreadable. Spec slice in TLA+; binary is still software with bugs. Coverage: Help Net Security. Repo: github.com/dredozubov/hazmat.

Pair it with last week’s open-weight defaults if you are running local Qwen/GLM bulk overnight — containment is orthogonal to which weights you pin. For shelves next to the agent box: Raspberry Pi / small SBC experiments, external NVMe enclosure for disposable agent worktrees, and a basic hardware security key so the human path is stronger than the agent path.

Takeaway

1. Today: walk the Microsoft support cutover page if anyone on your team still depends on consumer Group Chat, Labs, podcasts, or Deep Research — those are not “maybe later,” they are the Aug 18 bin.
2. This week: smoke AgentRadio (or any true non-blocking peer channel) on one interdependent SWE-Atlas-style investigation before you buy another max-tier solo seat.
3. Always-on: stop running coding agents as you on the laptop that holds SSH keys — Hazmat-class boundaries are cheaper than incident theater.
4. Keep yesterday’s Flash/Grok pins and the open-weight bulk lane; none of today’s story deletes those routes.

The spaghetti this morning is product subtraction meeting multi-agent gossip protocols. Model scoreboards still matter. Timing of evidence between workers and which features still exist on the client matter more for the ticket you ship before lunch.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.