---
schema_version: "newruntime-topic-hub-v0.2"
type: "topic_hub"
stable_id: "topic_hub:agent-harness"
slug: "agent-harness"
title: "Agent Harness - New Runtime"
description: "A New Runtime topic hub collecting signals, patterns, field notes, and public sources about agent harness."
retrieval_nugget: "A New Runtime topic hub collecting signals, patterns, field notes, and public sources about agent harness. Agent Harness is tracked here as an evidence-linked topic, not as a static glossary entry. The page connects raw observations to pattern hypotheses, longer analysis, and public sources. Use it as the canonical landing page before drilling into individual records."
answer: ["Agent Harness is tracked here as an evidence-linked topic, not as a static glossary entry.","The page connects raw observations to pattern hypotheses, longer analysis, and public sources.","Use it as the canonical landing page before drilling into individual records."]
search_intents: ["agent harness","agent harness AI agents","agent harness software"]
status: "featured"
last_updated: "2026-08-03"
record_date: "2026-08-03"
date_kind: "last_updated"
counts: {"total":121,"signals":104,"patterns":0,"posts":17,"atlas":0,"sources":193}
routes: {"html":"https://newruntime.com/topics/agent-harness/","markdown":"https://newruntime.com/topics/agent-harness.md","json":"https://newruntime.com/topics/agent-harness.json"}
source_urls: ["https://addyosmani.com/blog/good-spec","https://agentation.com/","https://ai.google.dev/gemini-api/docs/gemini-3","https://aitmpl.com/component/hook/dangerous-command-blocker","https://alilleybrinker.com/mini/gas-town-decoded","https://americanbanker.com/news/citi-mandates-ai-prompt-training-for-most-employees","https://ando.so/","https://anthropic.com/engineering/AI-resistant-technical-evaluations","https://anthropic.com/engineering/claude-code-sandboxing","https://anthropic.com/engineering/effective-context-engineering-for-ai-agents","https://anthropic.com/engineering/managed-agents","https://anthropic.com/news/claude-code-on-the-web"]
top_sources: ["https://addyosmani.com/blog/good-spec","https://agentation.com/","https://ai.google.dev/gemini-api/docs/gemini-3","https://aitmpl.com/component/hook/dangerous-command-blocker","https://alilleybrinker.com/mini/gas-town-decoded","https://americanbanker.com/news/citi-mandates-ai-prompt-training-for-most-employees","https://ando.so/","https://anthropic.com/engineering/AI-resistant-technical-evaluations","https://anthropic.com/engineering/claude-code-sandboxing","https://anthropic.com/engineering/effective-context-engineering-for-ai-agents","https://anthropic.com/engineering/managed-agents","https://anthropic.com/news/claude-code-on-the-web"]
---

# Agent Harness - New Runtime

A New Runtime topic hub collecting signals, patterns, field notes, and public sources about agent harness.

## Retrieval answer

A New Runtime topic hub collecting signals, patterns, field notes, and public sources about agent harness. Agent Harness is tracked here as an evidence-linked topic, not as a static glossary entry. The page connects raw observations to pattern hypotheses, longer analysis, and public sources. Use it as the canonical landing page before drilling into individual records.

## Patterns


## Field Notes

- [A Software Factory Connects Agents Through Verified Outcomes](https://newruntime.com/posts/agentic-sdlc-software-factory-loop/): Augment and Warp describe team-level agent loops that move work from trigger and specification through implementation, verification, release, and measured improvement.
- [Agents Should Search, Fetch, And Browse As Separate Operations](https://newruntime.com/posts/browserbase-search-fetch-browser-routing/): Browserbase separates discovery, content retrieval, and browser interaction so research agents do not launch a full browser merely to obtain a list of URLs.
- [ChatGPT Cuts Repeated Work Across The Agent Stack](https://newruntime.com/posts/chatgpt-agent-loop-efficiency-stack/): ByteByteGo's OpenAI engineering walkthrough connects persistent sessions, stable prompt prefixes, deferred tools, delta tokenization, cache-aware routing, and split inference.
- [Claude Code Auto Mode Gates Actions Instead Of Explanations](https://newruntime.com/posts/claude-code-auto-mode-action-gate/): Claude Code Auto Mode combines an input injection probe with a two-stage action classifier, preserving autonomy while exposing an honest residual miss rate.
- [Cline Hooks Put Deterministic Rules Inside The Agent Loop](https://newruntime.com/posts/cline-hooks-agent-harness-guardrails/): Cline's plugin hooks show how an agent harness can journal every run and block dangerous tool calls without waiting for the model to choose a guardrail.
- [Containment Caps An Agent's Blast Radius](https://newruntime.com/posts/anthropic-agent-containment-blast-radius/): Anthropic's three runtime patterns show why hard filesystem, network, credential, and trust boundaries carry more security weight than repeated approval prompts.
- [Anthropic Moves Large Tool Libraries Out Of Context](https://newruntime.com/posts/anthropic-tool-search-programmatic-calls/): Anthropic's Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples separate discovery, orchestration, and usage guidance for agents with large tool libraries.
- [Arcee Turns Scientific Post-Training Into A Run Ledger](https://newruntime.com/posts/arcee-open-model-science-post-training/): Arcee's open-model science write-up shows a 21-run post-training loop around Trinity Mini, held-out scientific environments, trace review, and a promoted specialist adapter.
- [Genkit Adds Progressive Disclosure For Agent Skills](https://newruntime.com/posts/genkit-agent-skills-progressive-disclosure/): Genkit now loads Agent Skills through middleware that discovers SKILL.md metadata first and activates full instructions, references, and scripts only when needed.
- [Cline Turns Recursive Self-Improvement Into Harness Work](https://newruntime.com/posts/cline-recursive-self-improvement-coding-agent/): Cline's Terminal-Bench run is not a singularity story; it is a concrete loop where an agent reads traces, patches the harness, reruns evals, and hands a PR to humans.
- [OpenAI Shows Efficiency Is a Full-Stack Agent Problem](https://newruntime.com/posts/openai-gpt-5-6-efficiency-stack/): OpenAI's GPT-5.6 efficiency write-up connects model training, inference optimization, and the Codex/ChatGPT Work harness into one compounding cost-performance loop.
- [OpenAI's ARC-AGI-3 Jump Was a Harness Result](https://newruntime.com/posts/openai-arc-agi-settings-harness/): OpenAI's ARC-AGI-3 write-up shows why agent benchmarks measure the model plus the runtime harness: retained reasoning and compaction changed both score and token use.
- [Claude Opus 5 Makes Prompting a Runtime Contract](https://newruntime.com/posts/claude-opus-5-prompt-contracts/): Anthropic's Opus 5 prompting guide shows that stronger models can make old harness defaults wrong: verbosity, effort, verification, delegation, and thinking mode all become runtime controls.
- [LangChain Deep Agents Shrink the Harness Instead of Adding More Prompt](https://newruntime.com/posts/langchain-deep-agents-lean-harness/): Deep Agents v0.7.0b2 cuts default-agent input tokens by 65% and tool-description tokens by 43%, turning harness efficiency into a first-class agent metric.
- [Open Secure AI Alliance Turns the AI-Safety Fight Into a Stack Question](https://newruntime.com/posts/open-secure-ai-alliance-open-defense-stack/): NVIDIA's Open Secure AI Alliance reframes open models, harnesses, identity, safe formats, scanners, and disclosure as shared defensive infrastructure for AI agents.
- [Agent Harness Optimization Is Becoming an Outer-Loop Discipline](https://newruntime.com/posts/agent-harness-optimization-outer-loop/): The next useful automation target is not another agent response. It is the controlled loop that changes prompts, tools, context, and routing, then keeps only improvements that survive evaluation.
- [The Coding Harness Is Becoming Independent From the Model](https://newruntime.com/posts/coding-agents-multi-model-harness/): Practitioners are routing different models through coding-agent workflows, while production systems increasingly choose model and effort per role instead of per product.

## Recent Raw Signals

- 2026-07-28: [Claude Opus 5 Turns Prompting Into Harness Calibration](https://newruntime.com/signals/claude-opus-5-prompt-contracts/)
- 2026-06-26: [Anthropic: Goal-Scoped Agent Loops](https://newruntime.com/signals/anthropic-goal-scoped-agent-loops/)
- 2026-06-02: [Claude Code: Verification Bandwidth](https://newruntime.com/signals/claude-code-verification-bandwidth/)
- 2026-06-02: [GitHub / cursor/plugins: Verification Bandwidth](https://newruntime.com/signals/github-cursor-plugins-verification-bandwidth/)
- 2026-05-21: [Asteroid: Task-Specific Ai Interfaces](https://newruntime.com/signals/asteroid-task-specific-ai-interfaces/)
- 2026-05-21: [Walkinglabs / Learn Harness Engineering: Verification Bandwidth](https://newruntime.com/signals/walkinglabs-learn-harness-engineering-verification-bandwidth/)
- 2026-05-20: [Google / Gemini Models: Agent-Ready Software](https://newruntime.com/signals/google-gemini-models-agent-ready-software/)
- 2026-05-07: [Claude: Goal-Scoped Agent Loops](https://newruntime.com/signals/claude-goal-scoped-agent-loops-2307/)
- 2026-05-07: [SKILL.md Acts as a Behavior Loader, Not a Better Prompt](https://newruntime.com/signals/skill-md-acts-as-a-behavior-loader-not-a-better-prompt/)
- 2026-05-06: [Claude: Goal-Scoped Agent Loops](https://newruntime.com/signals/claude-goal-scoped-agent-loops/)
- 2026-05-06: [Stripe Built ProtoDash Around Its Own Design System](https://newruntime.com/signals/stripe-built-protodash-around-its-own-design-system/)
- 2026-04-30: [GitHub / NousResearch/hermes-agent: Generative Media Infrastructure](https://newruntime.com/signals/github-nousresearch-hermes-agent-generative-media-infrastructure/)
- 2026-04-30: [GitHub / stripe/link-cli: Task-Specific Ai Interfaces](https://newruntime.com/signals/github-stripe-link-cli-task-specific-ai-interfaces/)
- 2026-04-24: [Ando: Ai-Native Operating Models](https://newruntime.com/signals/ando-ai-native-operating-models/)
- 2026-04-24: [Google Cloud / Gemini Enterprise Agent Platform: Ai-Native Operating Models](https://newruntime.com/signals/google-cloud-gemini-enterprise-agent-platform-ai-native-operating-models/)
- 2026-04-24: [Image generation prompting Guide](https://newruntime.com/signals/image-generation-prompting-guide/)

## Public Sources

- https://addyosmani.com/blog/good-spec
- https://agentation.com/
- https://ai.google.dev/gemini-api/docs/gemini-3
- https://aitmpl.com/component/hook/dangerous-command-blocker
- https://alilleybrinker.com/mini/gas-town-decoded
- https://americanbanker.com/news/citi-mandates-ai-prompt-training-for-most-employees
- https://ando.so/
- https://anthropic.com/engineering/AI-resistant-technical-evaluations
- https://anthropic.com/engineering/claude-code-sandboxing
- https://anthropic.com/engineering/effective-context-engineering-for-ai-agents
- https://anthropic.com/engineering/managed-agents
- https://anthropic.com/news/claude-code-on-the-web