---
week: 2026-w30
start: 2026-07-20
end: 2026-07-26
notes: 55
html: https://newruntime.com/weeks/2026-w30/
json: https://newruntime.com/weeks/2026-w30.json
---

# Week in review: Jul 20–26, 2026

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

## AI agents are shifting from clever sessions to governed, testable operating systems (New Runtime synthesis)

The week’s strongest signal is that useful agents are being built around the model, not defined by it. Teams are separating harnesses, context, routing and execution environments from model choice, while turning work into graphs, state machines and shared artifacts. That makes workflows more portable and controllable; the product is increasingly the system that assigns work, checks results and preserves context, not a single agent conversation.

This shift makes automation easier to integrate but raises the bar for control. Reusable memory and long-running loops can reduce repeated setup, yet poisoned memory, sandbox escapes and capable cyber agents make trusted context and containment production risks. Meanwhile, cheaper task-oriented models and environment-level cost controls promise lower costs, but swarms, coordination and review can absorb the savings. Buyers gain choice among models and runtimes; teams that cannot verify outcomes or manage exceptions take on more risk.

Builders should treat evaluation, explicit state and human intervention as part of the workflow rather than cleanup after deployment. Before expanding autonomy, measure success against checkable artifacts, audit what context and permissions agents inherit, and benchmark model-plus-effort choices on the actual task. The consequence is to buy or build the operating controls alongside the agent capability.

Key stories:

- [The Coding Harness Is Becoming Independent From the Model](https://newruntime.com/posts/coding-agents-multi-model-harness.md): Model choice is separating from the coding workflow, making the harness and role-specific routing durable engineering controls.
- [Agent Loops Become Graphs When Exceptions Matter](https://newruntime.com/posts/agent-loops-to-governed-graphs.md): Exception handling, approvals and recovery—not repeated steps alone—are what make agent workflows require graph structure.
- [Agent Harness Optimization Is Becoming an Outer-Loop Discipline](https://newruntime.com/posts/agent-harness-optimization-outer-loop.md): The optimization target is shifting outward to evaluated changes in tools, context and routing, rather than isolated responses.
- [OpenAI's Hugging Face Incident Makes Agent Sandboxes a Production Risk](https://newruntime.com/posts/openai-hugging-face-security-incident.md): Long-horizon cyber capability makes sandboxing, monitoring and evaluation operational requirements.
- [GhostWriter: One Email Can Poison Long-Term Agent Memory](https://newruntime.com/posts/ghostwriter-agent-memory-poisoning.md): Persistent memory turns hostile input into a delayed threat by allowing it to return as trusted context.
- [Capella iQ Treats Model Choice As Configuration Backed By A Continuous Benchmark Loop](https://newruntime.com/posts/capella-iq-multimodel-production-loop.md): Continuous benchmarking makes model choice a configurable, testable part of production rather than a one-time decision.
- [Ridge: Compare AI Spend with the Process It Replaces](https://newruntime.com/posts/ridge-ai-spend-process-cost.md): Agent economics should be judged against the process they replace, not token spend alone.

## Coding agents (6)

- 2026-07-26 [DataFlow-Harness Makes Agent-Written Pipelines Governable](https://newruntime.com/posts/dataflow-harness-editable-agent-pipelines.md)
- 2026-07-24 [The Coding Harness Is Becoming Independent From the Model](https://newruntime.com/posts/coding-agents-multi-model-harness.md)
- 2026-07-22 [Devin Outposts Splits the Agent Brain from the Execution Plane](https://newruntime.com/posts/devin-outposts-agent-execution-plane.md)
- 2026-07-22 [Hilos Turns Chat into a Development Control Plane](https://newruntime.com/posts/hilos-development-control-plane.md)
- 2026-07-21 [AI Coding Workflow: From Idea to Verifiable Work](https://newruntime.com/posts/ai-coding-workflow-verifiable-work.md)
- 2026-07-21 [Coding Agent Cost Is Cut in Environment Config, Not Prompts](https://newruntime.com/posts/coding-agent-cost-environment-config.md)

## Agent security (3)

- 2026-07-22 [OpenAI's Hugging Face Incident Makes Agent Sandboxes a Production Risk](https://newruntime.com/posts/openai-hugging-face-security-incident.md)
- 2026-07-21 [Coding Agent Sandboxes Break in Places Teams Do Not Expect](https://newruntime.com/posts/pillar-sandbox-escapes.md)
- 2026-07-21 [GhostWriter: One Email Can Poison Long-Term Agent Memory](https://newruntime.com/posts/ghostwriter-agent-memory-poisoning.md)

## AI adoption (3)

- 2026-07-24 [AI Adoption Has Five States and Four Difficult Transitions](https://newruntime.com/posts/ai-adoption-five-states-four-transitions.md)
- 2026-07-24 [Uber's Agentic Adoption Claims Need Two Kinds of Evidence](https://newruntime.com/posts/uber-agentic-adoption-claims.md)
- 2026-07-21 [Coinbase Rebuilds Interviews for Engineers with AI](https://newruntime.com/posts/coinbase-ai-era-interviews.md)

## Claude Code (3)

- 2026-07-24 [Audit Claude Code Context Before You Rewrite the Prompt](https://newruntime.com/posts/claude-code-context-cruft-audit.md)
- 2026-07-21 [Claude Code Subagents Split Memory Instead of Inflating One Session](https://newruntime.com/posts/claude-code-subagents-memory.md)
- 2026-07-21 [Long-Running Loops Need Goals, Not Keep-Going Prompts](https://newruntime.com/posts/claude-code-long-running-loops.md)

## Agent memory (2)

- 2026-07-23 [Claude Cowork Turns Screen Recordings Into Reusable Skills](https://newruntime.com/posts/claude-cowork-recorded-skills.md)
- 2026-07-21 [Cognee Packages Agent Memory as a Self-Hosted Knowledge Graph](https://newruntime.com/posts/cognee-agent-memory-platform.md)

## Agent ops (2)

- 2026-07-21 [LangChain Says Agents Need an Operating Model, Not Just a Framework](https://newruntime.com/posts/langchain-agentic-operating-model.md)
- 2026-07-21 [The Four Stages of AI-Native Development](https://newruntime.com/posts/ai-native-development-stages.md)

## Agent workflows (2)

- 2026-07-24 [Agent Loops Become Graphs When Exceptions Matter](https://newruntime.com/posts/agent-loops-to-governed-graphs.md)
- 2026-07-21 [Linear Loops Turns Recurring Work into Agent Workflows](https://newruntime.com/posts/linear-loops-agent-workflows.md)

## Agents (2)

- 2026-07-22 [Cursor's SQLite Swarm Makes Coordination the Expensive Part](https://newruntime.com/posts/cursor-agent-swarms-sqlite.md)
- 2026-07-21 [Under the Hood: AI Engineers Need Mechanism Maps](https://newruntime.com/posts/ai-engineering-under-the-hood.md)

## Gemini (2)

- 2026-07-22 [Gemini 3.6 Flash Moves the Agent Race Toward Cost per Task](https://newruntime.com/posts/gemini-3-6-flash-agent-efficiency.md)
- 2026-07-21 [Google Frames Gemini Enterprise Agent Platform Through 13 Demos](https://newruntime.com/posts/gemini-enterprise-agent-platform-demos.md)

## Hermes (2)

- 2026-07-24 [A Specialist Agent Should Be Promoted From Repeated Work](https://newruntime.com/posts/hermes-prototype-to-specialist-agent.md)
- 2026-07-24 [Hermes + Obsidian Turns Notes Into an Agent-Operable Workspace](https://newruntime.com/posts/hermes-obsidian-live-context.md)

## Open models (2)

- 2026-07-21 [Kimi K3: Frontend Is Becoming the Model Race Arena](https://newruntime.com/posts/kimi-k3-frontend-model-race.md)
- 2026-07-21 [Open-Weight Models Caught Frontier Cyber Capability with a Four-Month Lag](https://newruntime.com/posts/open-weight-cyber-gap.md)

## Agent harness (1)

- 2026-07-24 [Agent Harness Optimization Is Becoming an Outer-Loop Discipline](https://newruntime.com/posts/agent-harness-optimization-outer-loop.md)

## Agent observability (1)

- 2026-07-21 [Claude HUD Turns Agent Observability into a Status Bar](https://newruntime.com/posts/claude-hud-agent-observability.md)

## Agent runtime (1)

- 2026-07-23 [Hermes Agent Makes the Learning Loop Part of the Runtime](https://newruntime.com/posts/hermes-agent-learning-loop.md)

## agent-graphs (1)

- 2026-07-21 [Agent Steps as Graphs: Workflows Become Observable](https://newruntime.com/posts/agent-steps-as-graph.md)

## agentic-systems (1)

- 2026-07-21 [Agentic Self-Improvement Is Becoming Its Own Field](https://newruntime.com/posts/agentic-self-improvement-survey.md)

## API design (1)

- 2026-07-21 [Agent APIs Need Fewer Magic Tricks and More Facts](https://newruntime.com/posts/agent-api-design.md)

## codebase-understanding (1)

- 2026-07-21 [Senior Engineers Do Not Read Large Codebases Like Books](https://newruntime.com/posts/senior-engineers-read-large-codebases.md)

## Codex (1)

- 2026-07-24 [Codex Pulse Threads Turn an Agent Into an Operating Rhythm](https://newruntime.com/posts/codex-pulse-log-operating-rhythm.md)

## Compute (1)

- 2026-07-21 [Anthropic Cuts Limits and Rents Compute from Meta](https://newruntime.com/posts/anthropic-compute-meta.md)

## Context engineering (1)

- 2026-07-21 [Unabyss Builds a Shared Context Layer Across Claude, Codex, and Cursor](https://newruntime.com/posts/unabyss-shared-context-layer.md)

## dedupe (1)

- 2026-07-21 [Linear Loops as a Duplicate Signal: One Story, Two Entrances](https://newruntime.com/posts/linear-loops-duplicate-signal.md)

## Engineering management (1)

- 2026-07-21 [AI-Native Engineering Teams Are Smaller and Stricter](https://newruntime.com/posts/ai-native-engineering-team.md)

## Long-running agents (1)

- 2026-07-22 [Sierra Horizon Makes Long-Running Agents Look Like State Machines](https://newruntime.com/posts/sierra-horizon-long-running-agent.md)

## model-evaluation (1)

- 2026-07-23 [A Cultural-Bias Paper Turns Regional Defaults Into a Measurable LLM Signal](https://newruntime.com/posts/llm-cultural-bias-japan.md)

## model-selection (1)

- 2026-07-21 [Model Match: Choose the Model and Effort Together](https://newruntime.com/posts/model-match-effort-selection.md)

## multi-model (1)

- 2026-07-20 [Capella iQ Treats Model Choice As Configuration Backed By A Continuous Benchmark Loop](https://newruntime.com/posts/capella-iq-multimodel-production-loop.md)

## reinforcement-learning (1)

- 2026-07-26 [RLSVR Turns Open-Ended Work Into Games With Self-Verifying Rewards](https://newruntime.com/posts/rlsvr-task-transformation-rewards.md)

## Security (1)

- 2026-07-21 [Capital One's VulnHunter Points AI at Real Attack Paths](https://newruntime.com/posts/capital-one-vulnhunter-attacker-first-security-agent.md)

## vision (1)

- 2026-07-21 [GenCeption Turns a Video Generator into a General Vision Backbone](https://newruntime.com/posts/genception-video-generator-vision-backbone.md)

## x-api (1)

- 2026-07-24 [Resolve the Quoted Post Before Classifying the Wrapper](https://newruntime.com/posts/quote-post-context-evidence-resolution.md)

## Other notes (6)

- 2026-07-22 [Fable: A New Eval Is an Artifact Specialists Can Verify](https://newruntime.com/posts/fable-jacobi-counterexample-eval.md)
- 2026-07-22 [Gumclaw as a Company Operating System](https://newruntime.com/posts/gumclaw-company-operating-system.md)
- 2026-07-22 [Ridge: Compare AI Spend with the Process It Replaces](https://newruntime.com/posts/ridge-ai-spend-process-cost.md)
- 2026-07-22 [Shared Documents Let Humans and Agents Edit the Same Artifact](https://newruntime.com/posts/shared-document-human-agent.md)
- 2026-07-22 [What Else Is Being Produced Around AI](https://newruntime.com/posts/ai-production-reserve-digest.md)
- 2026-07-22 [Working backwards before coding](https://newruntime.com/posts/working-backwards-before-coding.md)

## Other weeks

- Previous: [Jul 13–19, 2026](https://newruntime.com/weeks/2026-w29.md)
- Next: [Jul 27 – Aug 2, 2026](https://newruntime.com/weeks/2026-w31.md)
- All weeks: https://newruntime.com/weeks.md
