--- title: Meta's EvoHarness-RL trains agents to manage their own... description: Meta's EvoHarness-RL is a research system that trains a policy to manage agent coordination scaffolding — the harness that tracks belief, progress, and... canonical: https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta's EvoHarness-RL trains agents to manage their own coordination scaffolding | daily.dev og:description: Meta's EvoHarness-RL is a research system that trains a policy to manage agent coordination scaffolding — the harness that tracks belief, progress, and... og:url: https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0 og:image: https://api.daily.dev/og/posts/4AR0u7SF0.png og:image:alt: Meta's EvoHarness-RL trains agents to manage their own coordination scaffolding 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. # Meta's EvoHarness-RL trains agents to manage their own coordination scaffolding **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Meta's EvoHarness-RL is a research system that trains a policy to manage agent coordination scaffolding — the harness that tracks belief, progress, and experience during long-running tasks — rather than hand-authoring it. Training uses supervised fine-tuning followed by cost-aware GRPO reinforcement learning on a Qwen3-8B base model, achieving 96.9% on ALFWorld. Two emergent behaviors appear: harness annealing (frequent harness calls get absorbed into model weights over time) and harness evolution (state gets compressed rather than growing unboundedly). The key insight is that long-horizon agent performance is bottlenecked by knowing when to use available resources, not by having more of them. ## Content Agent harnesses — the scaffolding that coordinates memory, progress tracking, and tool use during long-running tasks — are still mostly written by hand. That makes them brittle and hard to tune, especially for tasks that unfold over many steps. New research from Meta takes a different approach. Instead of having humans author the harness, they train a policy to manage it. ## How it works EvoHarness-RL learns a harness policy offline, then deploys it to construct and update external harness state during runtime. The harness exposes three things the policy can act on: **Belief** (what the agent currently thinks is true), **Progress** (how far along the task is), and **Experience** (what's been learned so far). Training happens in two stages. First, supervised fine-tuning teaches the model the action space — when and how to read, write, and consolidate harness state. Then cost-aware GRPO (a reinforcement learning method) trains the model to use those actions selectively rather than constantly. The base model is Qwen3-8B, which reaches 96.9% on ALFWorld with this approach. ## Two behaviors that emerge from training Two interesting dynamics show up after training: - **Harness annealing**: Recurring patterns of harness use get absorbed into the model weights over time. The agent gradually shifts from calling the harness frequently to accessing it only when it actually helps. - **Harness evolution**: Progress updates and experience consolidation compress the workspace into a compact, task-relevant state rather than an ever-growing log. ## Why this matters The core finding is that long-horizon agents benefit more from a trainable coordination policy than from bigger tools or larger memory buffers. The bottleneck isn't capacity — it's knowing when to use what you have. That's a meaningful reframe. A lot of agent research focuses on giving models more: more context, more tools, more memory. This suggests the harder problem is teaching agents to manage what they already have. ## Questions this post answers ### What benchmark score does EvoHarness-RL achieve on ALFWorld and what base model does it use? EvoHarness-RL achieves 96.9% on ALFWorld using Qwen3-8B as the base model. Training proceeds in two stages: supervised fine-tuning to teach the harness action space, followed by cost-aware GRPO reinforcement learning to make harness access selective rather than constant. _Developers building long-horizon agents track emerging benchmarks and coordination approaches like this on daily.dev._ ### What is harness annealing in the context of AI agent training? Harness annealing is an emergent behavior where recurring patterns of harness use get absorbed into model weights over training. The agent gradually shifts from calling the external harness frequently to accessing it only when it genuinely helps, reducing unnecessary coordination overhead without explicit supervision. _Teams designing agent memory and coordination systems find related research surfaced on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 4 discussions and 24 comments across x (as of 2026-08-10). **TL;DR:** Commenters broadly find the core insight compelling — that trainable coordination scaffolding beats brute-force scaling — but a recurring concern is debuggability and explainability of learned harnesses in production and regulated environments. **Sentiment:** 55% positive · 30% mixed · 15% skeptical **The case for** - Trainable coordination policies that learn when to read, update, and consolidate state are seen as a more meaningful multiplier than larger context or more tools. - Harness annealing absorbing recurring patterns into model weights is viewed as a genuine efficiency unlock. - Automating hand-authored harnesses removes a brittle, non-transferable tuning step that currently dominates agent-pipeline maintenance. - The 96.9% ALFWorld result on a Qwen3-8B base is considered a solid benchmark outcome. **The pushback** - Learned harness policies risk being black boxes, making it hard to debug failures or explain agent decisions in regulated workflows. - Offline training raises concerns about overfitting to the training distribution and poor generalization to genuinely novel deployment tasks. - It's unclear how harness annealing handles tasks where the optimal access pattern shifts mid-run. - The cost metric for GRPO (wall time vs. token spend) is unspecified, leaving the optimization target ambiguous. **By community** - x (mixed): Broadly receptive to the core idea that coordination policy is the real bottleneck, but split on whether learned harnesses can be trusted in production given debuggability and generalization concerns. **Hottest debate:** Whether learned harness policies are viable in regulated or production settings where the hand-authored harness serves as an auditable, explainable control surface. **Open questions** - How does harness annealing handle tasks where the optimal access pattern changes mid-run — does the policy detect the shift or follow the learned schedule? - Does offline training risk overfitting to the training distribution, and how does the system handle genuinely novel tasks at deployment? - What exactly does cost-aware GRPO optimize against — wall time or token spend? - How can teams inspect or audit why a learned harness changed its own operating rules? **Highlights** > @omarsar0 Learned harnesses will win the benchmarks. But in a regulated workflow the hand-authored harness is the point: it's the control surface you can read, pin, and defend when someone asks why the agent did that. A harness that tunes itself is one more thing you can't explain later. > — [vikasmalpani on x](https://x.com/vikasmalpani/status/2086719071194915145) > @omarsar0 Hand-authored harnesses are the hidden ceiling. Most of the distance between an agent demo and something that survives a month in production is scaffolding someone tuned by feel, and none of it transfers to the next task. Learning the harness policy is a more interesting result > — [aneesmerchant on x](https://x.com/aneesmerchant/status/2086706249098494021) > @omarsar0 Interesting approach. But does offline learning risk overfitting to the training distribution? How do they handle genuinely novel tasks at deployment? > — [sofiaopal92 on x](https://x.com/sofiaopal92/status/2086577219866939777) > @omarsar0 how does annealing handle tasks where the right access pattern changes halfway through a run? does the policy detect the shift or just stick to the learned schedule? > — [bountyAIhunter on x · 1 points](https://x.com/bountyAIhunter/status/2086514483015356577) > @omarsar0 This flips the scaling story Bigger tools and bigger memory are brute force A trainable coordination policy that learns when to read, update, and consolidate is the real multiplier > — [isssa0x on x](https://x.com/isssa0x/status/2086557262131171474) **Source threads** - [x](https://x.com/omarsar0/status/2086509069762981896) · 0 points · 24 comments - [x](https://x.com/omarsar0/status/2086634214019993619) · 0 points · 0 comments - [x](https://x.com/omarsar0/status/2086691597966872576) · 0 points · 0 comments - [x](https://x.com/petergyang/status/2086571241549320259) · 0 points · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0) ```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":"Meta's EvoHarness-RL trains agents to manage their own coordination scaffolding","url":"https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0"},"datePublished":"2026-08-10T05:50:39.799Z","dateModified":"2026-08-10T11:51:50.966Z","description":"Meta's EvoHarness-RL is a research system that trains a policy to manage agent coordination scaffolding — the harness that tracks belief, progress, and...","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,reinforcement-learning","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Meta's EvoHarness-RL trains agents to manage their own coordination scaffolding"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-s-evoharness-rl-trains-agents-to-manage-their-own-coordination-scaffolding-4ar0u7sf0#faq","mainEntity":[{"@type":"Question","name":"What benchmark score does EvoHarness-RL achieve on ALFWorld and what base model does it use?","acceptedAnswer":{"@type":"Answer","text":"EvoHarness-RL achieves 96.9% on ALFWorld using Qwen3-8B as the base model. Training proceeds in two stages: supervised fine-tuning to teach the harness action space, followed by cost-aware GRPO reinforcement learning to make harness access selective rather than constant. Developers building long-horizon agents track emerging benchmarks and coordination approaches like this on daily.dev."}},{"@type":"Question","name":"What is harness annealing in the context of AI agent training?","acceptedAnswer":{"@type":"Answer","text":"Harness annealing is an emergent behavior where recurring patterns of harness use get absorbed into model weights over training. The agent gradually shifts from calling the external harness frequently to accessing it only when it genuinely helps, reducing unnecessary coordination overhead without explicit supervision. Teams designing agent memory and coordination systems find related research surfaced on daily.dev."}}]} ```