---
week: 2026-w35
start: 2026-08-24
end: 2026-08-30
notes: 25
html: https://newruntime.com/weeks/2026-w35/
json: https://newruntime.com/weeks/2026-w35.json
---

# Week in review: Aug 24–30, 2026

25 Field Notes, grouped by topic. Each links to its Markdown twin; replace .md with .json for structured fields.

## Agent work is shifting from prompted sessions to persistent, permissioned systems (New Runtime synthesis)

Several notes point to the same change: agents are becoming standing participants in workflows, not just tools people summon. Cursor’s event subscriptions, Copilot in team chat, and Vercel Agent’s incident context connect work signals to proposed changes; shared plugins and hosted MCP search make capabilities easier to reach across clients. That lowers integration friction, but raises the stakes of deciding what an agent can see and do.

The expansion is not a case for simply removing human boundaries. Enterprise-managed connector access, MCP’s focus on identity, and Vercel’s sandbox challenge all show that permissions and containment remain active design problems. Portable Computer makes escalation a user-controlled choice, while cyber-capable models put access control at the center of deployment. Builders gain reach and continuity; buyers inherit more consequential permission and isolation decisions.

Evaluation and trust are also becoming runtime concerns: confidential double-blind testing, integrity checks, and task-specific eval pipelines all resist treating a score or longer run as proof of reliability. Meanwhile, free, heavily cached usage can outrun provenance, and agent-generated contributions can overwhelm maintainer review. Teams should instrument permissions, isolation, and review capacity alongside agent performance, and avoid granting persistent access or trusting benchmark claims without evidence.

Key stories:

- [Cursor cloud agents monitor PRs, Slack threads and scheduled tasks](https://newruntime.com/posts/cursor-cloud-agents-monitor-prs-slack-threads-and-scheduled-tasks.md): Event subscriptions make a standing goal, rather than a single prompt, the unit of agent work.
- [GitHub Copilot moves into Slack and Microsoft Teams](https://newruntime.com/posts/github-copilot-moves-into-slack-and-microsoft-teams.md): Bringing coding agents into team chat links issue triage and code changes to where collaboration already happens.
- [Enterprise-managed auth narrows agent connector permissions](https://newruntime.com/posts/enterprise-managed-auth-narrows-agent-connector-permissions.md): Identity-provider groups shift connector access from individual grants toward centrally managed permissions.
- [Vercel opens a 1M dollar Sandbox escape challenge](https://newruntime.com/posts/vercel-opens-a-1m-dollar-sandbox-escape-challenge.md): The sandbox challenge underscores that a virtual machine alone may not contain an agent.
- [Double-Blind Evals Move Benchmark Integrity into the Runtime](https://newruntime.com/posts/deepmind-double-blind-frontier-model-evaluations.md): Confidential computing makes benchmark integrity a runtime property for both evaluators and model owners.
- [Perplexity Portable Computer Makes Cloud Escalation a User-Controlled Boundary](https://newruntime.com/posts/perplexity-portable-computer-local-cloud-boundary.md): User-approved cloud escalation shows how orchestration can retain an explicit boundary around private work.
- [OpenClaw's Viral Growth Turns Maintainer Capacity into Security Infrastructure](https://newruntime.com/posts/openclaw-viral-growth-maintainer-security-boundary.md): Agent-generated contribution volume turns maintainer review into a security and supply-chain constraint.

## Agents (11)

- 2026-08-30 [Codex plugins and ChatGPT Apple Messages integration](https://newruntime.com/posts/codex-plugins-and-chatgpt-apple-messages-integration.md)
- 2026-08-30 [Cursor cloud agents monitor PRs, Slack threads and scheduled tasks](https://newruntime.com/posts/cursor-cloud-agents-monitor-prs-slack-threads-and-scheduled-tasks.md)
- 2026-08-30 [Exa ships hosted MCP search across a dozen agent clients](https://newruntime.com/posts/exa-search-through-mcp-and-vercel-ai-gateway.md)
- 2026-08-30 [Firecrawl Developer Index for coding agents](https://newruntime.com/posts/firecrawl-developer-index-for-coding-agents.md)
- 2026-08-30 [GitHub Copilot moves into Slack and Microsoft Teams](https://newruntime.com/posts/github-copilot-moves-into-slack-and-microsoft-teams.md)
- 2026-08-30 [Google Antigravity custom agents](https://newruntime.com/posts/google-antigravity-custom-agents.md)
- 2026-08-30 [OpenAI Zero Data Retention for frontier models](https://newruntime.com/posts/openai-zero-data-retention-for-frontier-models.md)
- 2026-08-30 [Ox Alpha tops OpenCode usage data](https://newruntime.com/posts/ox-alpha-tops-opencode-usage-data.md)
- 2026-08-30 [Vercel Agent comes to Slack](https://newruntime.com/posts/vercel-agent-comes-to-slack.md)
- 2026-08-27 [A Skill Router Must Be Able to Say No](https://newruntime.com/posts/skill-router-must-say-no.md)
- 2026-08-27 [The Minimum Integrity Stack for an Agent Leaderboard](https://newruntime.com/posts/minimum-integrity-stack-agent-leaderboard.md)

## Agent security (5)

- 2026-08-30 [Enterprise-managed auth narrows agent connector permissions](https://newruntime.com/posts/enterprise-managed-auth-narrows-agent-connector-permissions.md)
- 2026-08-30 [The new MCP roadmap](https://newruntime.com/posts/the-new-mcp-roadmap.md)
- 2026-08-30 [Vercel opens a 1M dollar Sandbox escape challenge](https://newruntime.com/posts/vercel-opens-a-1m-dollar-sandbox-escape-challenge.md)
- 2026-08-29 [Claude Mythos 5 comes to Claude Security](https://newruntime.com/posts/claude-mythos-5-comes-to-claude-security.md)
- 2026-08-26 [VM Isolation Is Only One Boundary for Cyber-Capable Agents](https://newruntime.com/posts/vm-isolation-is-only-one-agent-containment-boundary.md)

## Agent evals (2)

- 2026-08-28 [Double-Blind Evals Move Benchmark Integrity into the Runtime](https://newruntime.com/posts/deepmind-double-blind-frontier-model-evaluations.md)
- 2026-08-27 [An Anti-Cheat Prompt Is Not a Security Control](https://newruntime.com/posts/anti-cheat-prompt-is-not-security-control.md)

## new-feature (2)

- 2026-08-27 [Hours Are Visible; Autonomous Weeks Are Not](https://newruntime.com/posts/hours-visible-autonomous-weeks-not.md)
- 2026-08-27 [Two ExtractBenches Measure Different Systems](https://newruntime.com/posts/two-extractbenches-different-systems.md)

## Agent runtime (1)

- 2026-08-28 [Perplexity Portable Computer Makes Cloud Escalation a User-Controlled Boundary](https://newruntime.com/posts/perplexity-portable-computer-local-cloud-boundary.md)

## Anthropic (1)

- 2026-08-28 [Anthropic's Model Hardware Standard Defines a Control Boundary for Hardware Agents](https://newruntime.com/posts/anthropic-model-hardware-standard-control-boundary.md)

## figure (1)

- 2026-08-28 [Figure Index Turns Physical Data Collection into Robotics Infrastructure](https://newruntime.com/posts/figure-index-physical-data-infrastructure.md)

## langchain (1)

- 2026-08-28 [LangChain Turns Agent Evals into a Spec-to-Task Production Pipeline](https://newruntime.com/posts/langchain-agent-environments-spec-to-task-pipeline.md)

## Open source (1)

- 2026-08-28 [OpenClaw's Viral Growth Turns Maintainer Capacity into Security Infrastructure](https://newruntime.com/posts/openclaw-viral-growth-maintainer-security-boundary.md)

## Other weeks

- Previous: [Aug 17–23, 2026](https://newruntime.com/weeks/2026-w34.md)
- Next: [Aug 31 – Sep 6, 2026](https://newruntime.com/weeks/2026-w36.md)
- All weeks: https://newruntime.com/weeks.md
