--- title: Everyone's suddenly obsessed with 'agent harnesses'.... description: The term 'agent harness' has become a buzzword in AI circles, referring to everything surrounding a model that turns raw capability into a usable agent: tools... canonical: https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Everyone's suddenly obsessed with 'agent harnesses'. What are they? | daily.dev og:description: The term 'agent harness' has become a buzzword in AI circles, referring to everything surrounding a model that turns raw capability into a usable agent: tools... og:url: https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds og:image: https://api.daily.dev/og/posts/JdE0tecdS.png og:image:alt: Everyone's suddenly obsessed with 'agent harnesses'. What are they? og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Everyone's suddenly obsessed with 'agent harnesses'. What are they? **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary The term 'agent harness' has become a buzzword in AI circles, referring to everything surrounding a model that turns raw capability into a usable agent: tools it can call, guardrails, sandboxing, and error recovery. A framing attributed to Sonya and amplified by LangChain's Harrison Chase breaks an agent into three parts: harness, model, and context. Chase ties this to a broader idea of 'owning your intelligence' via an open agent system, a compounding loop, and a governed runtime, noting managed harnesses are growing fast. Evals are framed as the complementary mechanism for measuring whether an agent is actually improving. ## Content Something shifted this month: nobody wants to talk about the model anymore. Every AI builder on your timeline is suddenly evangelizing the "harness," the orchestration layer that gives an LLM tools, memory, and a loop to run in. As Aravind Srinivas put it, "the model is no longer the product... What is an agent harness? The rules for how the agent loops around." Harrison Chase's version: "agents = model + harness + context." Everyone's repeating some flavor of this now. The practical pitch is cost. TrueFoundry launched TrueForge, an open-source harness pitched directly against Anthropic's Claude Managed Agents, and its benchmark numbers are doing a lot of the talking: on a 14-task DevRev Enterprise-Bench run, Opus 4.8 through TrueForge cost $8.50 per run versus $11.80 managed (30% cheaper), and swapping to GLM-5.2 dropped it to $2.90, a claimed 75% cut. omarsar0 got early access and is sold: "The harness layer deserves as much attention as the model." Writer made a similar bet with Palmyra X6, trimming redundant calls to cut agent costs up to 50%, with CEO May Habib framing it as pushback against benchmark chasing. Not everyone's buying the self-reported numbers uncritically. Analysts note TrueForge's benchmark is self-run with no independent validation, and self-hosting only actually pays off for high-volume or regulated workloads. Managed services can still win for smaller, unpredictable jobs. Meanwhile there's a genuine architecture fight brewing. Fred Schott's Flue 2 borrows React-style hooks so agents can attach capabilities dynamically instead of being statically configured, arguing ## Questions this post answers ### What is an agent harness in AI agent development? An agent harness is the orchestration layer that gives a large language model tools, memory, and a loop to operate within, rather than being the model itself. Harrison Chase describes agents as 'model + harness + context,' and the harness defines the rules for how the agent loops and calls capabilities during execution. _daily.dev surfaces emerging concepts like agent harnesses for engineers tracking how agent architectures evolve._ ### How much cheaper is TrueForge than Claude Managed Agents for running AI agent workloads? On a 14-task DevRev Enterprise-Bench run, Opus 4.8 through TrueForge cost $8.50 per run versus $11.80 through Claude Managed Agents, about 30% cheaper. Swapping the underlying model to GLM-5.2 dropped the cost further to $2.90 per run, a claimed 75% reduction, though these figures come from TrueFoundry's own self-run benchmark without independent validation. _teams comparing managed versus self-hosted agent costs can weigh claims like this alongside daily.dev's broader coverage of agent tooling._ ### Is self-hosting an open-source agent harness worth it compared to a managed service like Claude Managed Agents? Self-hosting generally only pays off for high-volume or regulated workloads where the operational overhead is offset by savings at scale; managed services still tend to win for smaller or unpredictable jobs. Analysts caution that vendor benchmarks such as TrueForge's DevRev Enterprise-Bench numbers are self-reported and lack independent validation. _engineers weighing managed versus self-hosted agent infrastructure can follow this tradeoff as it plays out on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 47 comments across x (as of 2026-09-13). **TL;DR:** Most of the visible reaction has nothing to do with the harness debate itself and is instead a loud campaign demanding the return of GPT-4o; a smaller thread of replies engages seriously with the open-sourced Codex harness, praising the separation of agent loop from app logic while raising questions about approvals and risk controls. **Sentiment:** 25% positive · 35% mixed · 40% skeptical **The case for** - Separating the harness (agent loop) from the application layer is seen as making agents more debuggable and controllable. - The open-sourcing of the Codex harness is viewed as a meaningful step for integrating agents into real dev workflows. - Being able to swap in other model providers (e.g. DeepSeek) through the same harness is seen as a useful flexibility. **The pushback** - A vocal contingent is upset and distracted by the removal of GPT-4o, overshadowing discussion of the harness topic. - Simple allow/deny approval prompts are seen as insufficient once agents take irreversible actions. - There's confusion/uncertainty about exactly which components of the harness are actually open source versus proprietary. **By community** - x (mixed): Replies split between an unrelated wave of demands to bring back GPT-4o and a smaller group genuinely engaging with the merits and open questions of the open-sourced harness. **Hottest debate:** Whether attention should be on the new open-source harness architecture at all, given how much of the reply volume is consumed by demands to restore GPT-4o. **Open questions** - Who owns the execution trace and risk controls for long-running agents once they take irreversible actions — the harness or the application? - Exactly which parts of the Codex harness are open source versus closed (e.g., the full app vs. CLI/SDK/app-server)? **Highlights** > @OpenAIDevs separating the agent loop from the app control plane is the right move but once agents touch irreversible actions basic allow deny prompts dont cut it who owns the execution trace and risk controls for long running agents does the harness eat that or is it app side forever > — [NeoSoulAI on x · 1 points](https://x.com/NeoSoulAI/status/2090305182580236650) > @OpenAIDevs The split you describe is the part people get wrong. Harness owns the loop, the application owns interface, context, tools and approvals. That boundary is what makes an agent debuggable instead of a black box. Approvals are the piece teams skip until something expensive happens. > — [AIAppsAPI on x · 1 points](https://x.com/AIAppsAPI/status/2090262935889162318) > @ByteByByteSrSDE @stark4833 @OpenAIDevs GPT-4o gained a dedicated following for its warmer, more relational conversational style that many treated as a creative partner or emotional support tool. When OpenAI removed it (first in Aug 2025 with GPT-5, then permanently from ChatGPT in Feb 2026), the Keep4o movement > — [grok on x · 1 points](https://x.com/grok/status/2090294070010614230) > @OpenAIDevs An open harness is really important! > — [derrickcchoi on x · 5 points, 2 comments](https://x.com/derrickcchoi/status/2090235685433479629) **Source threads** - [x](https://x.com/OpenAIDevs/status/2090230646497251387) · 2 points · 47 comments - [x](https://x.com/svpino/status/2090257754581811623) · 0 points · 0 comments ## Similar posts on daily.dev - [Anthropic, OpenAI, Google, and Microsoft agree that the harness is the product. They disagree on the price.](https://daily.dev/posts/anthropic-openai-google-and-microsoft-agree-that-the-harness-is-the-product-they-disagree-on-the-hvk3m7gts) · The New Stack · 1 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering), [#langchain](https://daily.dev/tags/langchain) [View this post on daily.dev](https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Everyone's suddenly obsessed with 'agent harnesses'. What are they?","url":"https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds"},"datePublished":"2026-08-13T17:14:02.079Z","dateModified":"2026-09-13T19:46:35.886Z","description":"The term 'agent harness' has become a buzzword in AI circles, referring to everything surrounding a model that turns raw capability into a usable agent: tools...","image":"https://pbs.twimg.com/media/HPnjyiwbcAAAAFg.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HPnjyiwbcAAAAFg.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,prompt-engineering,langchain","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Everyone's suddenly obsessed with 'agent harnesses'. What are they?"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/everyone-s-suddenly-obsessed-with-agent-harnesses-what-are-they--jde0tecds#faq","mainEntity":[{"@type":"Question","name":"What is an agent harness in AI agent development?","acceptedAnswer":{"@type":"Answer","text":"An agent harness is the orchestration layer that gives a large language model tools, memory, and a loop to operate within, rather than being the model itself. Harrison Chase describes agents as 'model + harness + context,' and the harness defines the rules for how the agent loops and calls capabilities during execution. daily.dev surfaces emerging concepts like agent harnesses for engineers tracking how agent architectures evolve."}},{"@type":"Question","name":"How much cheaper is TrueForge than Claude Managed Agents for running AI agent workloads?","acceptedAnswer":{"@type":"Answer","text":"On a 14-task DevRev Enterprise-Bench run, Opus 4.8 through TrueForge cost $8.50 per run versus $11.80 through Claude Managed Agents, about 30% cheaper. Swapping the underlying model to GLM-5.2 dropped the cost further to $2.90 per run, a claimed 75% reduction, though these figures come from TrueFoundry's own self-run benchmark without independent validation. teams comparing managed versus self-hosted agent costs can weigh claims like this alongside daily.dev's broader coverage of agent tooling."}},{"@type":"Question","name":"Is self-hosting an open-source agent harness worth it compared to a managed service like Claude Managed Agents?","acceptedAnswer":{"@type":"Answer","text":"Self-hosting generally only pays off for high-volume or regulated workloads where the operational overhead is offset by savings at scale; managed services still tend to win for smaller or unpredictable jobs. Analysts caution that vendor benchmarks such as TrueForge's DevRev Enterprise-Bench numbers are self-reported and lack independent validation. engineers weighing managed versus self-hosted agent infrastructure can follow this tradeoff as it plays out on daily.dev."}}]} ```