Seven days, four themes: two more models slotted into Copilot's now-routine cadence, a vendor explaining a safety mechanism in genuine detail, a reliability story that landed on the same day as a platform land-grab, and the clearest step yet from agents that wait to be asked toward agents that act on their own.
The throughline, if there is one: capability keeps compounding, and every week now also ships something about who is accountable for it.
Two more models, same pattern
Grok 4.6 arrived on 14 August pitched at long agentic runs; Gemini 3.7 Flash arrived on 13 August behind a policy whose name includes "Preview". Both follow the now-standard shape: off by default for Business and Enterprise, provider list pricing, gradual rollout. That is the sixth and seventh model change in Copilot this month.
Anthropic explained the watermark
On 14 August Anthropic published how Claude's text watermarking works: a statistical pattern hidden in word choice, invisible to readers, costing nothing in tokens or latency. Code is barely touched by it, because the mechanism needs interchangeable word options that code does not offer. Detection can prove involvement; it can never prove human authorship, and that asymmetry is the whole basis for using it responsibly.
17 August: an outage and a launch, on the same day
GitHub ran an incident from 13:40 to 21:15 UTC affecting Git operations, the API, Actions, Issues, Pull Requests, Webhooks and Copilot authentication — roughly 20% error rates on web and API traffic, 50% on archive downloads.
Cursor announced Origin the same day: hosted repositories, pull requests and bidirectional GitHub sync, in early beta on all paid plans. The timing is coincidence, not strategy, but it put two versions of the same question in front of every engineering team at once — how many companies does your ability to ship depend on, and did anyone decide that number?
Governance parity and incident-response precision
On 18 August, JetBrains finally got the enterprise managed settings that VS Code has had — plugin governance, MCP allowlists, OpenTelemetry configuration, bypass-permissions control — closing a gap that mattered more given how much reach agents gained this month. The same day, GitHub let credential revocation target one token type at a time instead of nuking all of a user's credentials at once, which changes what a security team can afford to do on a weak signal rather than waiting for certainty.
Cloud agents stopped waiting to be asked
Cursor closed the week by giving Cloud Agents Subscriptions — they now wake on pull request activity, Slack threads or scheduled tasks and act without a new prompt — plus a /goal command for objectives that span sessions, and isolated VMs for subagents. Agents that only acted when prompted are bounded by your attention; agents that act on subscriptions are bounded by something much larger and less visible. That is worth a policy decision, not just an opt-in click.
Everything else that shipped
Rounding out a dense week:
- 13-14 Aug — OAuth apps gained 8-hour access tokens with 6-month refresh tokens and up to 10 redirect URIs; wildcard redirect matching, silently on for existing single-URI apps, is worth checking today
- 13 Aug — Cursor shipped Builds: pre-prepared cloud agent environments, 3x faster to first token, free and automatic for new environments
- 13 Aug — GitHub halved the share of dependency-graph packages with no known license, from 45% to 24%, across 170 million packages
- 19 Aug — an organization-level code quality Trends tab, tracking open findings by repository over 7, 14 or 30 days
What to actually do
A short list, weighted toward the items with a real deadline or default attached.
- Check wildcard redirect matching on your OAuth apps and turn it off where you don't need it — still the cheapest fix on any list this month
- If you administer JetBrains alongside VS Code, audit the new managed settings; the plugin and MCP controls you set up two weeks ago did not reach it until 18 August
- Decide a policy for event-driven agent actions before enabling Subscriptions for real work, not after the first unattended commit lands
- Write your organization's disclosure policy for AI-assisted text now, while the watermark detector is still hypothetical rather than live