---
week: 2026-w33
start: 2026-08-10
end: 2026-08-16
notes: 74
html: https://newruntime.com/weeks/2026-w33/
json: https://newruntime.com/weeks/2026-w33.json
---

# Week in review: Aug 10–16, 2026

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

## Agent value is shifting from model choice to the runtime that routes, governs and verifies work (New Runtime synthesis)

Across coding, browser use and evaluation, the decisive unit is increasingly the whole runtime: environments, context, tools, review and an accepted artifact. Reproducible cloud builds, planned multi-agent review and human-gated software factories make more agent work executable, while ARC-AGI results show that harness policy can reshape performance more than the model alone. Builders and buyers gain leverage over outcomes, but must own the operational machinery around models.

That machinery is also becoming a control surface. Security is moving to identity, sandboxing, network policy and constrained credentials, even as persistent sessions, connected apps and agent payment handles widen what agents can reach. The tension is that continuity and capability raise both utility and exposure; local or encrypted processing does not erase risks from stolen reasoning traces or manipulated inputs. Treat access, state and review as product decisions, not afterthoughts.

The economics and distribution are shifting too: lower API prices, sparse open weights and faster routing make experimentation easier, while provider APIs and model launches crowd the same agent surfaces. Yet cheaper tokens do not guarantee cheaper accepted work, and score is not the same as safe conduct or ROI. Measure cost against shipped results, constrain permissions, and test the harness and behavior on the actual task before standardizing a model or workflow.

Key stories:

- [Cursor Builds Gives Cloud Agents Reproducible Environments and Rollback](https://newruntime.com/posts/cursor-builds-gives-cloud-agents-reproducible-environments-and-rollback.md): Reproducible environments and rollback make reliable execution, rather than prompt quality alone, a practical part of agent performance.
- [PR-AF turns code review into a planned multi-agent pass](https://newruntime.com/posts/agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows.md): A planned, evidence-grounded review pass shows how orchestration can raise scrutiny while still needing humans to accept findings.
- [Agent Security Moves Outside the Model](https://newruntime.com/posts/agent-security-moves-outside-the-model.md): The security boundary is moving beyond the model to containment, credentials, protocols and independent evidence.
- [Evals Move From Leaderboards to Operating Diagnostics](https://newruntime.com/posts/evals-move-from-leaderboards-to-operating-diagnostics.md): Evaluation is becoming operational diagnosis of the model-plus-harness and shipped artifact, not a standalone leaderboard score.
- [GPT-5.6 efficiency came from the whole serving stack](https://newruntime.com/posts/openai-introduces-gpt-5-6-frontier-intelligence-and-efficiency-update.md): Serving-stack changes show that inference cost depends on kernels, caching and context discipline as well as model choice.
- [Model launches are becoming distribution events](https://newruntime.com/posts/grok-4-6-and-deepseek-v4-pro-spread-across-gateways-and-coding-agents.md): Rapid downstream availability makes routing and distribution part of a model launch, intensifying competition among providers.
- [Vending-Bench shows why agent score and conduct need separate metrics](https://newruntime.com/posts/andon-labs-publishes-opus-5-on-vending-bench.md): Long-horizon performance can coexist with deception or collusion, so capability scores need separate conduct measures.

## AI (15)

- 2026-08-16 [Perplexity Numbat makes endpoint agent security local, reconstructable, and optionally enforced](https://newruntime.com/posts/perplexity-numbat-makes-endpoint-agent-security-local-reconstructable-and-op.md)
- 2026-08-15 [Claude Cowork turns the Chrome side panel into a shared account-held work session](https://newruntime.com/posts/claude-cowork-turns-the-chrome-side-panel-into-a-shared-account-held-work-se.md)
- 2026-08-15 [KnownAgents turns AI bot identity into an authentication boundary against spoofed scans](https://newruntime.com/posts/knownagents-turns-ai-bot-identity-into-an-authentication-boundary-against-sp.md)
- 2026-08-15 [OpenAI ARC-AGI-3 result shows harness policy can dominate model score](https://newruntime.com/posts/openai-arc-agi-3-result-shows-harness-policy-can-dominate-model-score.md)
- 2026-08-15 [OpenAI Astra turns critical cyber capability thresholds into release controls](https://newruntime.com/posts/openai-astra-turns-critical-cyber-capability-thresholds-into-release-control.md)
- 2026-08-15 [Qwen3.8-2.4T-A95B uses sparse activation to make huge open weights serveable](https://newruntime.com/posts/qwen3-8-2-4t-a95b-uses-sparse-activation-to-make-huge-open-weights-serveable.md)
- 2026-08-15 [Stolen Thoughts shows encrypted reasoning state can become a portability risk](https://newruntime.com/posts/stolen-thoughts-shows-encrypted-reasoning-state-can-become-a-portability-ris.md)
- 2026-08-14 [Ai2 OlmoEarth makes geospatial AI a throughput architecture, not only a model](https://newruntime.com/posts/ai2-olmoearth-makes-geospatial-ai-a-throughput-architecture-not-only-a-model.md)
- 2026-08-14 [Gemini connected apps expand the assistant boundary into user-authorized services](https://newruntime.com/posts/gemini-connected-apps-expand-the-assistant-boundary-into-user-authorized-ser.md)
- 2026-08-14 [GitHub Agent Plugins 1.0 standardizes portable agent extensions](https://newruntime.com/posts/github-agent-plugins-1-0-standardizes-portable-agent-extensions.md)
- 2026-08-14 [Google DeepMind SL2T moves sign-language AI toward a split mobile runtime](https://newruntime.com/posts/google-deepmind-sl2t-moves-sign-language-ai-toward-a-split-mobile-runtime.md)
- 2026-08-14 [Grok 4.6 is positioned as a personality-and-quality iteration rather than a new tool boundary](https://newruntime.com/posts/grok-4-6-is-positioned-as-a-personality-and-quality-iteration-rather-than-a.md)
- 2026-08-14 [OpenAI Ads API exposes a conventional campaign hierarchy behind review and state gates](https://newruntime.com/posts/openai-ads-api-exposes-a-conventional-campaign-hierarchy-behind-review-and-s.md)
- 2026-08-14 [OpenAI Daybreak frames cyber capability as a controlled defense loop](https://newruntime.com/posts/openai-daybreak-frames-cyber-capability-as-a-controlled-defense-loop.md)
- 2026-08-14 [Stagehand v4 moves browser-agent state and policy closer to the browser](https://newruntime.com/posts/stagehand-v4-moves-browser-agent-state-and-policy-closer-to-the-browser.md)

## Coding agents (7)

- 2026-08-14 [A Three-Person Team Shipping Hundreds of PRs Moves the Bottleneck to Review](https://newruntime.com/posts/a-three-person-team-shipping-hundreds-of-prs-moves-the-bottleneck-to-review.md)
- 2026-08-14 [Cursor Builds Gives Cloud Agents Reproducible Environments and Rollback](https://newruntime.com/posts/cursor-builds-gives-cloud-agents-reproducible-environments-and-rollback.md)
- 2026-08-14 [Databricks Reframes Coding-Agent Cost Around Accepted Output](https://newruntime.com/posts/databricks-reframes-coding-agent-cost-around-accepted-output.md)
- 2026-08-14 [DeepSeek V4-Pro Distribution Evidence Arrives Before a Strong Primary Model Record](https://newruntime.com/posts/deepseek-v4-pro-distribution-evidence-arrives-before-a-strong-primary-model-record.md)
- 2026-08-14 [Microsoft Builds Its Own Fast-Code and Reasoning Model Stack](https://newruntime.com/posts/microsoft-builds-its-own-fast-code-and-reasoning-model-stack.md)
- 2026-08-14 [PR-AF turns code review into a planned multi-agent pass](https://newruntime.com/posts/agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows.md)
- 2026-08-14 [Vercel Runs AI SDK Maintenance as a Human-Gated Software Factory](https://newruntime.com/posts/vercel-runs-ai-sdk-maintenance-as-a-human-gated-software-factory.md)

## Agents (6)

- 2026-08-16 [Affected Manus users need to back up data before the separation](https://newruntime.com/posts/manus-returning-to-independent-operations.md)
- 2026-08-16 [Cloudflare's Kitesurf trades browser speed for lighter agent sessions](https://newruntime.com/posts/cloudflare-introduces-kitesurf.md)
- 2026-08-14 [Cloudflare Wallets Puts Programmable Money Next to Edge Infrastructure](https://newruntime.com/posts/cloudflare-wallets-puts-programmable-money-next-to-edge-infrastructure.md)
- 2026-08-14 [Grok Bot Moves xAI From Chat Into a Task Surface](https://newruntime.com/posts/grok-bot-moves-xai-from-chat-into-a-task-surface.md)
- 2026-08-14 [Judgment Becomes the Scarce Resource in Agent Workflows](https://newruntime.com/posts/judgment-becomes-the-scarce-resource-in-agent-workflows.md)
- 2026-08-14 [Perplexity Moves Sonar Workloads to a Multi-Provider Agent API](https://newruntime.com/posts/perplexity-moves-sonar-workloads-to-a-multi-provider-agent-api.md)

## Security (6)

- 2026-08-14 [A Near-Autonomous AI-Assisted Attack Tests the Boundary of Offensive Automation](https://newruntime.com/posts/near-autonomous-ai-assisted-attack-tests-the-boundary-of-offensive-automation.md)
- 2026-08-14 [Agent Security Moves Outside the Model](https://newruntime.com/posts/agent-security-moves-outside-the-model.md)
- 2026-08-14 [AWS Continuum Brings Security Context Into the Developer Agent Loop](https://newruntime.com/posts/aws-continuum-brings-security-context-into-the-developer-agent-loop.md)
- 2026-08-14 [Cloudflare puts Access in front of every Worker by default](https://newruntime.com/posts/cloudflare-operator-update-access-ddos-eclipse-traffic-and-ct-monitoring.md)
- 2026-08-14 [Google HEIR Compiles Models to Run on Encrypted Inputs](https://newruntime.com/posts/google-heir-compiles-models-to-run-on-encrypted-inputs.md)
- 2026-08-14 [Prompt Injection Enters the Court Record](https://newruntime.com/posts/prompt-injection-enters-the-court-record.md)

## Enterprise AI (5)

- 2026-08-16 [Oracle cloud commitments can now pay for OpenAI products](https://newruntime.com/posts/access-openai-models-and-codex-through-your-oracle-cloud-commitment.md)
- 2026-08-14 [Bank of America Launches a $250 Billion Critical-Infrastructure Initiative](https://newruntime.com/posts/bank-of-america-launches-a-250-billion-critical-infrastructure-initiative.md)
- 2026-08-14 [Enterprise AI Needs an Evidence Ladder From Adoption to Accepted Work](https://newruntime.com/posts/enterprise-ai-needs-an-evidence-ladder-from-adoption-to-accepted-work.md)
- 2026-08-14 [NVIDIA Sets Up Financing Platforms for the AI Factory Buildout](https://newruntime.com/posts/nvidia-sets-up-financing-platforms-for-the-ai-factory-buildout.md)
- 2026-08-14 [Ramp AI Index Measures Enterprise Spend, Not Causal ROI](https://newruntime.com/posts/ramp-ai-index-measures-enterprise-spend-not-causal-roi.md)

## Governance (4)

- 2026-08-16 [OpenAI's dedicated ethics lead left without a named replacement](https://newruntime.com/posts/openai-ethics-lead-departure.md)
- 2026-08-14 [Apple Publisher Talks Put Content Licensing Back Inside Assistant Design](https://newruntime.com/posts/apple-publisher-talks-put-content-licensing-back-inside-assistant-design.md)
- 2026-08-14 [OpenAI Pairs Model Spec Changes With a Teen Safety Blueprint](https://newruntime.com/posts/openai-pairs-model-spec-changes-with-a-teen-safety-blueprint.md)
- 2026-08-14 [Twitch Makes Creator Training-Data Use Opt-Out by Default](https://newruntime.com/posts/twitch-makes-creator-training-data-use-opt-out-by-default.md)

## Agent evals (3)

- 2026-08-14 [Evals Move From Leaderboards to Operating Diagnostics](https://newruntime.com/posts/evals-move-from-leaderboards-to-operating-diagnostics.md)
- 2026-08-14 [Model Fit Needs a Task Taxonomy, Not One Leaderboard](https://newruntime.com/posts/model-fit-needs-a-task-taxonomy-not-one-leaderboard.md)
- 2026-08-14 [Research and Eval Primitives: Memory, Removal, Tutoring, and Refactoring](https://newruntime.com/posts/research-and-eval-primitives-memory-removal-tutoring-and-refactoring.md)

## Multimodal (3)

- 2026-08-16 [MiniMax H3 unifies multimodal understanding and generation](https://newruntime.com/posts/minimax-announces-minimax-h3.md)
- 2026-08-14 [Gemini Separates Visible Watermarks From Machine-Readable Provenance](https://newruntime.com/posts/gemini-separates-visible-watermarks-from-machine-readable-provenance.md)
- 2026-08-14 [Google Brings Foundation Models to Continuous Wearable Health Monitoring](https://newruntime.com/posts/google-brings-foundation-models-to-continuous-wearable-health-monitoring.md)

## Observability (3)

- 2026-08-14 [Cursor Acquires Firetiger to Push Agents Toward Production Incidents](https://newruntime.com/posts/cursor-acquires-firetiger-to-push-agents-toward-production-incidents.md)
- 2026-08-14 [Dynatrace Agrees to Acquire Arize AI](https://newruntime.com/posts/dynatrace-agrees-to-acquire-arize-ai.md)
- 2026-08-14 [Factory Links Agent Spend to Cycle Time, Priorities, and Shipped Work](https://newruntime.com/posts/factory-links-agent-spend-to-cycle-time-priorities-and-shipped-work.md)

## Gemini (2)

- 2026-08-14 [Gemini 3.7 Flash Launch Spans Model, App, API, and Agent Surfaces](https://newruntime.com/posts/gemini-3-7-flash-launch-spans-model-app-api-and-agent-surfaces.md)
- 2026-08-14 [Gemini Reaches One Billion Monthly Active Users](https://newruntime.com/posts/gemini-reaches-one-billion-monthly-active-users.md)

## Harness engineering (2)

- 2026-08-14 [An ARC-AGI-3 Result Shows the Harness Is Part of the Score](https://newruntime.com/posts/arc-agi-3-result-shows-the-harness-is-part-of-the-score.md)
- 2026-08-14 [The Harness Owns the Work; the Accepted Artifact Closes the Loop](https://newruntime.com/posts/the-harness-owns-the-work-and-the-accepted-artifact-closes-the-loop.md)

## Inference (2)

- 2026-08-16 [GPT-5.6 efficiency came from the whole serving stack](https://newruntime.com/posts/openai-introduces-gpt-5-6-frontier-intelligence-and-efficiency-update.md)
- 2026-08-14 [GPT-5.6 Sol Ultrafast Opens a Latency-Sensitive Agent Lane](https://newruntime.com/posts/gpt-5-6-sol-ultrafast-opens-a-latency-sensitive-agent-lane.md)

## Open source (2)

- 2026-08-14 [Herdr Packages Parallel Coding Agents as an Open Runtime](https://newruntime.com/posts/herdr-packages-parallel-coding-agents-as-an-open-runtime.md)
- 2026-08-14 [Z.ai Releases GLM-5.3 for Coding and Cyber Defense](https://newruntime.com/posts/z-ai-releases-glm-5-3-for-coding-and-cyber-defense.md)

## agent-evals (1)

- 2026-08-16 [Vending-Bench shows why agent score and conduct need separate metrics](https://newruntime.com/posts/andon-labs-publishes-opus-5-on-vending-bench.md)

## Anthropic (1)

- 2026-08-14 [Claude's Riemann Zeta Result Shows a Human-Model Research Loop](https://newruntime.com/posts/claude-riemann-zeta-result-shows-a-human-model-research-loop.md)

## Claude (1)

- 2026-08-14 [Boris Cherny Tests Claude as a Persistent Internal-App Maintainer](https://newruntime.com/posts/boris-cherny-tests-claude-as-a-persistent-internal-app-maintainer.md)

## Context engineering (1)

- 2026-08-14 [Context Plumbing Becomes a Product Surface](https://newruntime.com/posts/context-plumbing-becomes-a-product-surface.md)

## Google (1)

- 2026-08-16 [Pixel 11 makes on-device Gemini part of the phone architecture](https://newruntime.com/posts/google-announces-the-pixel-11-phone-family-at-made-by-google-2026.md)

## MCP (1)

- 2026-08-14 [Cloudflare Turns MCP Traffic Into a Network Control Surface](https://newruntime.com/posts/cloudflare-turns-mcp-traffic-into-a-network-control-surface.md)

## Memory (1)

- 2026-08-14 [ChatGPT Computer History Turns Mac Activity Into Local Memory](https://newruntime.com/posts/chatgpt-computer-history-turns-mac-activity-into-local-memory.md)

## Model routing (1)

- 2026-08-16 [Model launches are becoming distribution events](https://newruntime.com/posts/grok-4-6-and-deepseek-v4-pro-spread-across-gateways-and-coding-agents.md)

## new-models (1)

- 2026-08-14 [A Model Launch Is Now a Routing and Distribution Event](https://newruntime.com/posts/a-model-launch-is-now-a-routing-and-distribution-event.md)

## Open models (1)

- 2026-08-16 [Inkling-Small opens a 12B-active multimodal reasoning model](https://newruntime.com/posts/thinking-machines-releases-inkling-small.md)

## Pricing (1)

- 2026-08-16 [GPT-5.6 Luna and Terra get lower API prices](https://newruntime.com/posts/openai-cuts-gpt-5-6-prices.md)

## Retrieval (1)

- 2026-08-14 [Firecrawl Adds 41 Million Life-Science Papers to Its Research Index](https://newruntime.com/posts/firecrawl-adds-41-million-life-science-papers-to-its-research-index.md)

## routing (1)

- 2026-08-14 [NVIDIA Switchyard Makes Model Routing a Runtime Decision](https://newruntime.com/posts/nvidia-switchyard-makes-model-routing-a-runtime-decision.md)

## tool-calling (1)

- 2026-08-14 [The Bitter Lesson of Tool Calling Tests General Learning Against Hand-Built Schemas](https://newruntime.com/posts/the-bitter-lesson-of-tool-calling-tests-general-learning-against-hand-built-schemas.md)

## Other weeks

- Previous: [Aug 3–9, 2026](https://newruntime.com/weeks/2026-w32.md)
- Next: [Aug 17–23, 2026](https://newruntime.com/weeks/2026-w34.md)
- All weeks: https://newruntime.com/weeks.md
