{"schema_version":"newruntime-week-v0.1","type":"week_in_review","generated_at":"2026-11-18T18:00:02.626Z","key":"2026-w32","label":"Aug 3–9, 2026","start":"2026-08-03","end":"2026-08-09","count":85,"routes":{"html":"https://newruntime.com/weeks/2026-w32/","json":"https://newruntime.com/weeks/2026-w32.json","markdown":"https://newruntime.com/weeks/2026-w32.md","index":"https://newruntime.com/weeks.json","previous":"https://newruntime.com/weeks/2026-w31.json","next":"https://newruntime.com/weeks/2026-w33.json"},"note_routes":{"html":"https://newruntime.com/posts/{slug}/","markdown":"https://newruntime.com/posts/{slug}.md","json":"https://newruntime.com/posts/{slug}.json"},"summary":{"label":"New Runtime synthesis","headline":"AI agents are moving from clever sessions to governed, measurable operating systems","paragraphs":["The central shift is from treating agents as chat-driven helpers to building durable work systems around them. Coding workflows now emphasize task framing, isolated workspaces, persistent state, review, and verified completion; organization agents add scoped identity, permissions, memory, and approval gates. Across evaluation, observability, and cost notes, the same lesson recurs: the useful unit is a completed job that can be checked, not an impressive run or cheap token.","That operational discipline comes with a tension: more autonomy makes containment and authority boundaries more consequential. AISI’s incident points to permissive authorization, not a sandbox escape, while security notes argue that action gates, scoped credentials, and hard runtime limits matter more than repeated prompts. Durable memory and shared skills also make work easier to resume and reuse, but raise portability, ownership, and lifecycle questions. The gains accrue to teams that can govern and verify workflows; ad hoc users inherit more failure and lock-in risk.","This continues the prior movement toward agent infrastructure, but this week pushes it into organizational routines: PM, legal, payroll, field service, and fund administration are becoming reviewable workflows rather than blank chat surfaces. Cheaper coding and more accessible local runtimes widen who can build, while a crowded field of models and tools makes selection harder. Builders and buyers should compare systems on replayable task outcomes, permissions, and exportability, and track whether promised savings survive retries, review, and rework."],"key_stories":[{"slug":"coding-agent-economics-completed-task","title":"Coding-Agent Economics Moves From Token Spend To Cost Per Completed Task","why":"Cost shifts from token totals to verified task completion, including the retries and review that determine real economics."},{"slug":"anthropic-cyber-eval-incident-review","title":"AISI's Cyber Incident Was An Authorization Failure, Not A Sandbox Escape","why":"The cyber incident makes enforced authority boundaries—not merely containment claims—the central safety question."},{"slug":"portable-agent-session-test","title":"The Agent Lock-In Test Is A Full Session Export","why":"A full session export turns portability and deletion into concrete tests of agent lock-in."},{"slug":"supabase-agent-evals-runtime-contract","title":"Supabase Evals Separates The Task, Runtime, Agent, And Scorer","why":"Separating task, runtime, agent, and scorer makes comparisons reproducible rather than anecdotal."},{"slug":"anthropic-agent-containment-blast-radius","title":"Containment Caps An Agent's Blast Radius","why":"Hard limits on filesystem, network, and credentials cap blast radius more directly than approval prompts."},{"slug":"llm-costs-operational-risk-controls","title":"LLM Costs Need Prevention, Detection, And Mitigation","why":"Volatile spend becomes an operational risk requiring prevention, detection, and mitigation across agent loops."},{"slug":"model-routing-needs-eval-observability-loop","title":"Model Routing Is A Control Loop, Not A Static Cost Switch","why":"Routing only saves money when traces, outcome evaluations, and fallback policy connect model choice to quality."}],"source_note_count":85,"generated_at":"2026-10-27T04:04:17.249Z"},"topics":[{"slug":"coding-agents","label":"Coding agents","count":15,"notes":[{"slug":"senior-ai-coding-workflow-as-loop","title":"AI Coding Workflow Is Becoming A Senior Review Loop","day":"2026-08-05"},{"slug":"pm-workflows-move-into-coding-agents","title":"PM Work Is Moving Into Coding-Agent Surfaces","day":"2026-08-05"},{"slug":"phone-agent-control-plane","title":"A Phone Is A Control Plane For Persistent Agents","day":"2026-08-04"},{"slug":"agents-md-constitution-after-60b-tokens","title":"AGENTS.md Becomes The Operating Constitution","day":"2026-08-04"},{"slug":"cloudflare-codex-astro-agent-factory","title":"Cloudflare Shows The Governance And Execution Rails Of A Software Factory","day":"2026-08-04"},{"slug":"coding-agent-economics-completed-task","title":"Coding-Agent Economics Moves From Token Spend To Cost Per Completed Task","day":"2026-08-04"},{"slug":"mirrorcode-program-reconstruction-benchmark","title":"MirrorCode Tests Whether An Agent Can Rebuild A Whole Program From Behavior","day":"2026-08-04"},{"slug":"orca-worktree-integration-control-plane","title":"Orca Solves Agent Workspace Isolation, Then Surfaces The Integration Work","day":"2026-08-04"},{"slug":"port22-mobile-agent-control-plane","title":"port22 Turns The Phone Into A Live Control Plane For Coding Agents","day":"2026-08-04"},{"slug":"ramp-swebench-production-coding-eval","title":"Ramp SWE-Bench Points Coding Evals Back At Production Work","day":"2026-08-04"},{"slug":"warp-agent-cli-persistent-terminal-runtime","title":"Warp Agent CLI Makes The Terminal Session Part Of The Agent Runtime","day":"2026-08-04"},{"slug":"agentic-sdlc-software-factory-loop","title":"A Software Factory Connects Agents Through Verified Outcomes","day":"2026-08-03"},{"slug":"cline-hooks-agent-harness-guardrails","title":"Cline Hooks Put Deterministic Rules Inside The Agent Loop","day":"2026-08-03"},{"slug":"cursor-google-workspace-agent-loop","title":"Cursor Turns Google Workspace Into Agent Context","day":"2026-08-03"},{"slug":"evocode-bench-multi-turn-regressions","title":"EvoCode-Bench Exposes Multi-Turn Regression Risk","day":"2026-08-03"}]},{"slug":"agents","label":"Agents","count":10,"notes":[{"slug":"organization-agent-runtime-qm-vercel-eve","title":"An Organization Agent Is A Scoped Runtime, Not One Slack Bot","day":"2026-08-05"},{"slug":"cloudflare-computer-harness-in-app","title":"Cloudflare Computer Lets The Agent Loop Choose Its Runtime","day":"2026-08-05"},{"slug":"comp-ai-agent-first-crm","title":"Comp AI Treats CRM As An Agent's Evidence Notebook","day":"2026-08-05"},{"slug":"anthropic-cyber-eval-incident-review","title":"AISI's Cyber Incident Was An Authorization Failure, Not A Sandbox Escape","day":"2026-08-04"},{"slug":"aisi-unsanctioned-agent-actions","title":"AISI's Incident Was An Authorization Failure, Not A Sandbox Escape","day":"2026-08-04"},{"slug":"browserbase-search-fetch-browser-routing","title":"Agents Should Search, Fetch, And Browse As Separate Operations","day":"2026-08-03"},{"slug":"amazon-quick-agentic-catalog-context-boundary","title":"Amazon Quick Makes Catalog Semantics The Agent Boundary","day":"2026-08-03"},{"slug":"bcg-agentic-field-service-operating-loop","title":"Field Service Agents Need An Operating Loop, Not A Chat Window","day":"2026-08-03"},{"slug":"gpt-live-continuous-media-loop","title":"GPT-Live Separates The Conversation Loop From Deep Reasoning","day":"2026-08-03"},{"slug":"gusto-cofounder-blank-canvas-workflow","title":"Gusto Solves The Agent Blank Canvas With Scheduled Work","day":"2026-08-03"}]},{"slug":"open-models","label":"Open models","count":5,"notes":[{"slug":"kimi-k3-long-horizon-coding-model","title":"Kimi K3 Keeps Long-Context Coding In The Model Race","day":"2026-08-05"},{"slug":"cerebras-moe-router-gradient-null-expert","title":"A Balanced MoE Router Can Still Be Functionally Dead","day":"2026-08-03"},{"slug":"interconnects-model-artifacts-adoption-dashboard","title":"Interconnects Turns Model Releases Into A Longitudinal Adoption Dataset","day":"2026-08-03"},{"slug":"sequoia-open-model-dependency-paradox","title":"Open Weights Can Still Create A Strategic Dependency","day":"2026-08-03"},{"slug":"fireworks-lora-fullft-three-test-protocol","title":"Run Three Tests Before Replacing LoRA With Full Fine-Tuning","day":"2026-08-03"}]},{"slug":"agent-memory","label":"Agent memory","count":3,"notes":[{"slug":"mem0-dream-background-memory-compaction","title":"Mem0 Dream Adds Background Compaction To Agent Memory","day":"2026-08-05"},{"slug":"graphify-local-knowledge-graph","title":"Graphify Makes Agent Memory Inspectable Before It Is Compressed","day":"2026-08-04"},{"slug":"abbel-belief-state-memory","title":"ABBEL Treats Memory as an Explicit Belief State","day":"2026-08-03"}]},{"slug":"agent-observability","label":"Agent observability","count":3,"notes":[{"slug":"opik-agent-diagnostics-trace-layer","title":"Opik Turns Agent Traces Into a Debugging Surface","day":"2026-08-03"},{"slug":"digibee-opik-prompt-versioning-loop","title":"Prompt Versioning Is Becoming Agent Operations","day":"2026-08-03"},{"slug":"observable-job-agent-opik-langgraph","title":"The Observable Job Agent Is a Useful Agent-Product Template","day":"2026-08-03"}]},{"slug":"inference","label":"Inference","count":3,"notes":[{"slug":"llamafile-0105-local-model-runtime-refresh","title":"Llamafile 0.10.5 Refreshes The Local Model Runtime Around New Model Shapes","day":"2026-08-05"},{"slug":"meta-mslk-genai-kernel-library","title":"Meta MSLK Moves GenAI Performance Work Into A Reusable Kernel Library","day":"2026-08-04"},{"slug":"chatgpt-agent-loop-efficiency-stack","title":"ChatGPT Cuts Repeated Work Across The Agent Stack","day":"2026-08-03"}]},{"slug":"evals","label":"Agent evals","count":2,"notes":[{"slug":"supabase-agent-evals-runtime-contract","title":"Supabase Evals Separates The Task, Runtime, Agent, And Scorer","day":"2026-08-05"},{"slug":"smevals-small-model-evaluation-framework","title":"smevals Treats Small-Model Testing As A First-Class Agent Harness","day":"2026-08-04"}]},{"slug":"agent-harness","label":"Agent harness","count":2,"notes":[{"slug":"canva-founder-loop-agent-harness","title":"Canva's Founder Loop Looks Like A Human Agent Harness","day":"2026-08-04"},{"slug":"kiro-unified-agent-harness","title":"Kiro Moves Agent Logic Below The IDE, CLI, Web, And Mobile Surfaces","day":"2026-08-03"}]},{"slug":"agent-security","label":"Agent security","count":2,"notes":[{"slug":"claude-code-auto-mode-action-gate","title":"Claude Code Auto Mode Gates Actions Instead Of Explanations","day":"2026-08-03"},{"slug":"anthropic-agent-containment-blast-radius","title":"Containment Caps An Agent's Blast Radius","day":"2026-08-03"}]},{"slug":"codex","label":"Codex","count":2,"notes":[{"slug":"agentmicro-menu-bar-codex-observability","title":"AgentMicro Makes Parallel Codex Work Visible Without A Cloud Queue","day":"2026-08-04"},{"slug":"education-agent-plugins-and-vibe-coding-course","title":"OpenAI And Google Package Agent Learning Around Role-Specific Workflows","day":"2026-08-04"}]},{"slug":"developer-tools","label":"Developer tools","count":2,"notes":[{"slug":"simon-llm-032-eventful-tool-runtime","title":"LLM 0.32 Turns Tool Calls Into A Stored, Resumable Event Stream","day":"2026-08-04"},{"slug":"qm-multiplayer-agent-harness","title":"QM Turns Team Agents Into Scoped Runtime State","day":"2026-08-04"}]},{"slug":"governance","label":"Governance","count":2,"notes":[{"slug":"agent-skills-at-organizational-scale","title":"Hundreds Of Agent Skills Create A Lifecycle Problem, Not A Prompt Library","day":"2026-08-04"},{"slug":"juniper-square-fay-admin-oversight-agent","title":"Juniper Square Fay Turns Fund Admin Into A Reviewable Agent Workflow","day":"2026-08-04"}]},{"slug":"knowledge-work","label":"Knowledge work","count":2,"notes":[{"slug":"uber-first-pass-prd-reviewer","title":"Uber Makes PRD Review A First-Pass AI Workflow","day":"2026-08-05"},{"slug":"github-legal-copilot-cli-workflows","title":"GitHub Legal Shows How Non-Engineers Can Build Reviewable CLI Workflows","day":"2026-08-04"}]},{"slug":"mcp","label":"MCP","count":2,"notes":[{"slug":"workos-management-mcp-agent-admin","title":"WorkOS Management MCP Makes SaaS Admin State Agent-Addressable","day":"2026-08-04"},{"slug":"copilotkit-react-mcp-client-interface","title":"CopilotKit Brings MCP Tool Calls Into The React Interface","day":"2026-08-03"}]},{"slug":"model-routing","label":"Model routing","count":2,"notes":[{"slug":"luna-worker-model-routing","title":"Luna Worker Routing Needs A Written Contract","day":"2026-08-04"},{"slug":"model-routing-needs-eval-observability-loop","title":"Model Routing Is A Control Loop, Not A Static Cost Switch","day":"2026-08-04"}]},{"slug":"models","label":"Models","count":2,"notes":[{"slug":"deepseek-v4-flash-codex-budget-model","title":"DeepSeek V4-Flash Belongs In The Harness, Not The Hype Loop","day":"2026-08-04"},{"slug":"qwen38-max-active-parameter-economics","title":"Qwen3.8-Max Makes Active Parameters The Useful Question","day":"2026-08-04"}]},{"slug":"multimodal","label":"Multimodal","count":2,"notes":[{"slug":"shieldstral-inference-time-safety-policy","title":"Shieldstral Turns Guardrails Into An Inference-Time Policy","day":"2026-08-04"},{"slug":"video-use-transcript-first-editing","title":"video-use Treats The Transcript As The Agent's Primary Video Interface","day":"2026-08-04"}]},{"slug":"open-source","label":"Open source","count":2,"notes":[{"slug":"open-source-devtools-personal-software","title":"Coding Agents Make The Personal Devtool Fork Practical","day":"2026-08-05"},{"slug":"genoffice-ai-native-office-suite","title":"GenOffice Moves The Agent Loop Into The Office Engine","day":"2026-08-05"}]},{"slug":"agent-economics","label":"Agent economics","count":1,"notes":[{"slug":"llm-costs-operational-risk-controls","title":"LLM Costs Need Prevention, Detection, And Mitigation","day":"2026-08-03"}]},{"slug":"agent-identity","label":"Agent identity","count":1,"notes":[{"slug":"chaindrop-agent-identity-control-plane","title":"ChainDrop Shows Why Agent Security Starts With Identity And Release Authority","day":"2026-08-04"}]},{"slug":"skills","label":"Agent skills","count":1,"notes":[{"slug":"org-skills-as-operating-memory","title":"Claude Skills Point Toward Org-Level Operating Memory","day":"2026-08-05"}]},{"slug":"agent-workflows","label":"Agent workflows","count":1,"notes":[{"slug":"formula1-aws-agentic-data-onboarding","title":"Formula 1 And AWS Turn Data-Source Onboarding Into A Three-PR Agent Workflow","day":"2026-08-04"}]},{"slug":"ai-infrastructure","label":"AI infrastructure","count":1,"notes":[{"slug":"portable-agent-session-test","title":"The Agent Lock-In Test Is A Full Session Export","day":"2026-08-05"}]},{"slug":"ai-hardware","label":"ai-hardware","count":1,"notes":[{"slug":"openai-apple-trade-secrets-dispute","title":"OpenAI And Apple Turn Agent Hardware Recruiting Into A Trade-Secrets Test","day":"2026-08-03"}]},{"slug":"async-work","label":"async-work","count":1,"notes":[{"slug":"async-agent-supervision-conductor-chatgpt-activity","title":"Async Agents Need Both Workspaces And An Attention Queue","day":"2026-08-05"}]},{"slug":"benchmarks","label":"benchmarks","count":1,"notes":[{"slug":"mercor-apex-accounting-productivity-benchmark","title":"Mercor APEX-Accounting Measures AI Productivity Inside A Real Profession","day":"2026-08-04"}]},{"slug":"continual-learning","label":"continual-learning","count":1,"notes":[{"slug":"continual-adaptation-self-distillation-loop","title":"Macaron And On-Policy Self-Distillation Point Toward Post-Deployment Adaptation Loops","day":"2026-08-03"}]},{"slug":"education","label":"education","count":1,"notes":[{"slug":"college-educator-reviewable-workflow","title":"The College Educator Plugin Encodes A Reviewable Vertical Workflow","day":"2026-08-04"}]},{"slug":"enterprise-ai","label":"Enterprise AI","count":1,"notes":[{"slug":"aws-bedrock-native-web-search-tool","title":"Amazon Bedrock Adds Web Search As A Server-Side Grounding Tool","day":"2026-08-04"}]},{"slug":"generative-video","label":"generative-video","count":1,"notes":[{"slug":"minimax-h3-apple-silicon-local-video","title":"MiniMax H3 Makes Local Video Generation A Hardware And Licensing Story","day":"2026-08-04"}]},{"slug":"human-preference","label":"human-preference","count":1,"notes":[{"slug":"human-preference-model-routing-layer","title":"Design Arena Turns Human Preference Into A Data Layer For Model Selection","day":"2026-08-03"}]},{"slug":"inference-optimization","label":"inference-optimization","count":1,"notes":[{"slug":"final-bench-regression-proof-optimization","title":"FINAL-Bench Counts Speedups Only When The Model Still Matches The Quality Contract","day":"2026-08-04"}]},{"slug":"infrastructure","label":"Infrastructure","count":1,"notes":[{"slug":"google-session-aware-load-balancing-ai-agents","title":"Session-Aware Load Balancing Treats Long-Lived AI Work As A Commitment","day":"2026-08-04"}]},{"slug":"local-inference","label":"local-inference","count":1,"notes":[{"slug":"local-inference-hardware-thresholds","title":"Local Inference Is Crossing More Hardware Tiers, But Configuration Still Defines The Claim","day":"2026-08-04"}]},{"slug":"long-running-agents","label":"Long-running agents","count":1,"notes":[{"slug":"kiro-crew-long-running-agent-work","title":"Kiro Crew Adds A Long-Running Work Layer Above The Shared Agent Harness","day":"2026-08-04"}]},{"slug":"loop-engineering","label":"Loop engineering","count":1,"notes":[{"slug":"google-adk-loop-engineering-state-pruning","title":"Loop Engineering Needs State Pruning, Not Infinite Chat","day":"2026-08-03"}]},{"slug":"memory","label":"Memory","count":1,"notes":[{"slug":"contextual-agent-memory-four-layer-system","title":"A Vector Store Is Not An Agent Memory System","day":"2026-08-03"}]},{"slug":"multi-agent","label":"Multi-agent","count":1,"notes":[{"slug":"google-agent-teamwork-file-mediated-crews","title":"Agent Teams Collaborate Through Files, Gates, and Roles","day":"2026-08-03"}]},{"slug":"product-management","label":"product-management","count":1,"notes":[{"slug":"ant-murphy-outcome-led-product-management","title":"Ant Murphy Argues Cheap Coding Moves Product Work Beyond Epics","day":"2026-08-04"}]},{"slug":"vertical-agents","label":"vertical-agents","count":1,"notes":[{"slug":"ellis-private-credit-agentic-book","title":"Ellis Frames Private Credit Agents Around One Source-Verifiable Book","day":"2026-08-04"}]}]}
