Four developments this week complete a different kind of picture than last week's wave of GA launches and pricing moves. The focus has shifted from what models can do to the infrastructure layer around them: how teams govern agent work, where agents run cheaply, who owns the compute, and what happens when geopolitical rules override a closed deal. Each of these is now a concrete decision for operators, not a planning assumption.
Slack Code puts AI coding agents in the group chat
On 21 August, Salesforce launched Slack Code: dedicated Slack channels where engineering teams can observe, steer, and approve AI coding agents in real time. Four agents are supported at launch — Anthropic's Claude Code, Cognition's Devin, GitHub Copilot, and Vercel's v0 agent. The product is available on all Slack plans from day one; users supply their own agent credentials.
The mechanics matter more than the brand. Inside a Code channel, the agent's plan is pinned and visible to everyone. Team members can post feedback mid-task, review code diffs in place, preview HTML output before it ships, and approve changes without leaving the channel. When a task closes, the channel archives automatically and preserves a complete audit trail. There is no separate tool to open, no terminal window to share, no private agent session running outside the team's line of sight.
The governance implication is underappreciated. Solo vibe-coding is ungovernable: one developer, one agent, no record. Code channels make the agent's reasoning and output observable to peers, product owners, and reviewers from the first prompt. That changes the compliance posture for any team that needs to explain what was built and why. Slack's interim CEO Rob Seaman framed it as making code no longer the bottleneck — the more precise observation is that it makes the agent's work auditable without slowing it down.
Cloudflare Kitesurf: a browser built for agents, not people
Earlier this month — and worth surfacing now as browser-use agents move into production — Cloudflare launched Kitesurf on 6 August: a web browser runtime purpose-built for AI agents that runs entirely in V8 isolates on Cloudflare Workers. It uses 3 to 7 times less CPU and memory than Chromium for common agentic tasks, passes more than 235,000 Web Platform Tests (97 percent DOM coverage, 96 percent HTML), and exposes a Chrome DevTools Protocol endpoint compatible with existing Puppeteer, Playwright, and MCP clients.
The architecture strips out everything a human browser requires — tabs, themes, extensions, pixel-perfect rendering — and focuses on what an agent actually needs: a stateless DOM, reliable form interaction, tab management, and low-overhead per-session compute. It was built in Rust and WebAssembly on top of Firefox's Stylo CSS parser and the Boa ECMAScript engine, all inside Cloudflare's existing Workers isolate infrastructure.
For operators running browser-use agents at any volume, the cost case is direct. Chromium-based browser sessions are memory-intensive and expensive to scale. An agent fleet that runs against Kitesurf sits on infrastructure that is substantially cheaper per session at the same output fidelity, without requiring any change to existing Playwright or Puppeteer automation code. Available free while in beta.
Mistral pre-sells a gigawatt of sovereign European compute it has not yet built
On 11 August, Mistral announced it intends to reach 200 megawatts of European AI compute capacity by end of 2027 and a full gigawatt by 2030, and that it is financing the build by selling the capacity forward before it exists. The instrument is called a European Compute Unit: a pre-purchase contract against future capacity, underwritten by early enterprise commitments from ASML, CMA CGM, and Amadeus.
The numbers give the ambition scale. Research firm Epoch AI calculates that a one-gigawatt AI data centre requires roughly $38 billion in capital expenditure, with server and GPU procurement consuming the largest share. Mistral currently operates less than 200 megawatts, spread across a 44-megawatt facility near Paris that became operational in the second quarter of 2026, a 23-megawatt site in Sweden built with EcoDataCenter, and a 10-megawatt facility in Les Ulis, France.
The strategic logic is aimed directly at regulated enterprise buyers. European firms navigating GDPR, the EU AI Act, and emerging sector-specific data-localisation rules face a jurisdictional question that US-hosted model APIs cannot fully answer: where does the inference actually run, and whose law governs it? Mistral is selling that answer as a vertically integrated product. If the compute build lands on schedule, it will have replaced the standard model-as-a-service pitch with something more durable: a European AI stack from model weights to physical silicon, with provable sovereignty at every layer.
China completes the Meta-Manus unwind — and starts deleting data today
The data-deletion window in the China-mandated unwind of Meta's $2 billion Manus acquisition opened this morning. The backup deadline was 7:59 a.m. Singapore time on 23 August; data created by certain users on or after 29 December 2025 — the acquisition close date — is being deleted between 8:00 a.m. SGT today and 24 August.
The background: Meta completed the roughly $2 billion Manus acquisition on 29 December 2025. China's National Development and Reform Commission issued an unwinding directive in April 2026 under foreign investment and technology-export rules, ordering the parties to reverse the transaction. Manus, incorporated in Singapore, has since resumed independent operations and is migrating its user base out of Meta's infrastructure.
The operational signal for any company building AI tools that touch user or training data is stark. A closed, paid acquisition can be unwound after the fact under Chinese data rules. Unwinding can include compulsory deletion of production data created after the deal closed. That risk does not require the acquirer to be a Chinese entity, nor the target to be based in China — Manus is Singaporean, Meta is American. What it requires is that the deal or the data flow in some way engages Chinese data-export jurisdiction. Any operator evaluating acquisitions, partnerships, or data agreements that route through China-regulated entities should treat this as a live precedent, not a one-off edge case.
The through-line this week is that the infrastructure question and the governance question have merged. Whether the issue is audit trails for agent output (Slack Code), compute cost at scale (Kitesurf), jurisdictional clarity for inference (Mistral), or data sovereignty in a deal structure (Manus), the same underlying question keeps surfacing: under whose rules does this AI work operate, and what happens when those rules assert themselves? Operators who have not mapped an explicit answer to that question are now running on borrowed time.