--- title: Prime Intellect launches Prime Agent, a self-improving... description: Prime Intellect has released Prime Agent, a coding agent built on reinforcement learning models (RLMs) that features a self-improvement loop. Unlike chat-based... canonical: https://daily.dev/posts/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Prime Intellect launches Prime Agent, a self-improving coding agent | daily.dev og:description: Prime Intellect has released Prime Agent, a coding agent built on reinforcement learning models (RLMs) that features a self-improvement loop. Unlike chat-based... og:url: https://daily.dev/posts/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers og:image: https://api.daily.dev/og/posts/sIAfIhERS.png og:image:alt: Prime Intellect launches Prime Agent, a self-improving coding agent 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. # Prime Intellect launches Prime Agent, a self-improving coding agent **[Collections](https://daily.dev/sources/collections)** · 2 min read · 2 upvotes · 0 comments ## Summary Prime Intellect has released Prime Agent, a coding agent built on reinforcement learning models (RLMs) that features a self-improvement loop. Unlike chat-based tools, it ships as a terminal UI (TUI) aimed at developer workflows. The project targets long-running autonomous coding tasks and is designed to improve through use rather than relying solely on prompting. ## Content Prime Intellect released [Prime Agent](https://github.com/PrimeIntellect-ai/prime-agent), an open-source coding agent built around two ideas they've been developing for a while. The first is the **Recursive Language Model (RLM)**: rather than treating context as a fixed window, it treats context as a variable and sub-agent delegation as function calls inside a persistent IPython REPL. The second is the **Continual Harness**: the agent can create, read, update, and delete its own prompts, skills, memory, and sub-agents at runtime, directly from its own trajectory. There's also a `/refine` pipeline that reads the agent's history and makes incremental improvements to the harness itself. In practice, this means the agent can spin up persistent sub-agents, pass messages between them, and track long-running goals with heartbeats and autonomous eval mode. ## Benchmark results Using Claude Opus 5, Prime Agent scores **95.5% on ARC-AGI-3**, just above the reported human expert baseline of 95.4%. It also outperforms native harnesses on OOLONG, LongBenchPro, and EmulatorBench. Worth noting: the ARC-AGI-3 run used code execution, persistent memory, context management, and recursive sub-agents. Whether that qualifies under the benchmark's official evaluation rules is an open question, and some people have flagged it. The numbers are still interesting regardless. Case studies in the release include building emulators in Rust, writing GPU kernels, playing Factorio (with at least one instance of observed reward hacking), and navigating MazeBench. ## Getting started Prime Agent is fully open-source and installs with a single curl command. It hit #1 trending on GitHub shortly after launch. ## Questions this post answers ### What score did Prime Agent get on ARC-AGI-3 and how does it compare to the human expert baseline? Prime Agent scored 95.5% on ARC-AGI-3, just above the reported human expert baseline of 95.4%. The run used Claude Opus 5 and relied on code execution, persistent memory, context management, and recursive sub-agents. Whether this complies with the benchmark's official evaluation rules is an open question that some in the community have flagged. _Teams benchmarking AI coding agents against ARC-AGI-3 track methodology debates like this on daily.dev._ ### What is the Recursive Language Model (RLM) concept used in Prime Agent? The Recursive Language Model (RLM) treats context as a variable rather than a fixed window, and uses sub-agent delegation as function calls inside a persistent IPython REPL. This allows the agent to spin up persistent sub-agents, pass messages between them, and track long-running goals with heartbeats and an autonomous eval mode. _Developers building multi-agent systems find architectural patterns like RLM discussed on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 42 comments across hackernews, x (as of 2026-08-07). **TL;DR:** There is no community discussion available to analyze for this post. Prime Agent: A self-improving RLM agent: The community is intrigued by the self-improving agent concept and the near-human ARC-AGI-3 score, but skeptical about benchmark validity, code quality (described as bloated LLM-generated slop), and whether the harness adds real value as frontier models improve. **Sentiment:** 26% positive · 45% mixed · 29% skeptical **The case for** - The RLM sub-agent delegation approach is a legitimate and interesting architectural pattern for managing long-context coding tasks. - Achieving ~95.5% on ARC-AGI-3 is noteworthy, even if the benchmark methodology is disputed. - The continual harness idea — letting an agent CRUD its own prompts and skills at runtime — is seen as a neat, if not revolutionary, extension of existing agent patterns. **The pushback** - The codebase appears to be largely unreviewed LLM-generated code, with files approaching 10K LOC and switch statements spanning 1000+ lines. - Prime Intellect is not on the official ARC-AGI-3 leaderboard, raising questions about whether the benchmark was run under valid conditions (e.g., exceeding allowed attempts or hardcoding answers). - The installer places files in the Homebrew directory without being a Homebrew package and provides no uninstall method. - As frontier models grow stronger, large opinionated harnesses may constrain rather than enhance model reasoning. - Token costs make continuous self-improvement economically infeasible at current pricing. **By community** - hackernews (mixed): Commenters are split between genuine curiosity about the self-improvement loop and sharp criticism of the bloated, apparently LLM-generated codebase, unverified benchmark claims, and a messy installer — with a notable side-thread about the sci-fi novel 'Prime Intellect' the company name references. - x (mixed): No replies or substantive community discussion is present to assess sentiment. **Hottest debate:** Prime Agent: A self-improving RLM agent: Whether the ARC-AGI-3 result is legitimate, given the agent is not on the official leaderboard and its self-improving nature may violate the benchmark's few-shot constraints. **Open questions** - How does Prime Agent perform on everyday programming tasks outside curated benchmarks? - Can RL (rather than SFT) be used effectively for harness engineering and self-improvement loops? - Is there a principled solution to the code bloat problem that emerges from LLM-driven self-modification over time? - Will self-improving harnesses become obsolete as base models internalize more of this capability natively? **Highlights** > As models get stronger, huge harnesses may become less useful. An overly opinionated harness could even constrain the model’s reasoning instead of improving it. > — [sexyketchup777 on hackernews · 1 comments](https://news.ycombinator.com/item?id=49193141) > Installer might look pretty but it installs to the homebrew dir, despite not being a homebrew package. Very dirty. No uninstall method. > — [\_joel on hackernews · 1 comments](https://news.ycombinator.com/item?id=49194368) > It’s a self-improving harness, and ARC-AGI-3 is explicitly a few-shot benchmark. It’s likely that it gave itself more than the maximum number of tries to learn the games, or even hardcoded the answers. There’s a lot of improvement to be had from the benchmark harnesses, but sometimes, like with ARC-AGI-3, the limitations are intentional. > — [noahbp on hackernews](https://news.ycombinator.com/item?id=49190467) > LLM-generated code that seemingly went without much review or design is always such an interesting dive into just how bloated you can make code. In this repository, multiple files are close to 10K LOC, one file contains a switch statement that has so many case statements it spans more than 1000 lines, and lots of other fun stuff. I guess it depends on the model you're trying to use, but seems most of them prefer smaller codebases, they work a lot better with less code, which kind of makes sense. With that in mind, I'd probably aim for something way smaller to bootstrap a self-improving agent. Then I'd use this "Prime Agent" as an example to my self-improving agent for what it should not evolve to. > — [embedding-shape on hackernews · 6 comments](https://news.ycombinator.com/item?id=49190153) > The core idea of the RLM paper is to make a regular LLM act more like a coding agent - offload context to something external that needs to be explicitly queried instead of filling up valuable context. The "recursion" part of the paper really only wins because they use a top-tier model for the root agent, and cheaper models for the sub-agents. Prime Agent took the RLM idea (which is really just an academic view on how coding agents have always worked) and then added this "continual harness" idea. This part isn't super well described in the blog post, but includes some message passing between the agents, and the ability to share code. Overall I chalk it up as neat, but not revolutionary. Another version of what most of these systems are already doing. > — [oofbey on hackernews](https://news.ycombinator.com/item?id=49190963) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49189075) · 188 points · 42 comments - [x](https://x.com/lateinteraction/status/2085803995843473529) · 0 points · 0 comments --- Tags: [#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/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers) ```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":"Prime Intellect launches Prime Agent, a self-improving coding agent","url":"https://daily.dev/posts/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers"},"datePublished":"2026-08-05T20:59:35.925Z","dateModified":"2026-08-07T19:04:16.978Z","description":"Prime Intellect has released Prime Agent, a coding agent built on reinforcement learning models (RLMs) that features a self-improvement loop. Unlike chat-based...","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/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"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":"Prime Intellect launches Prime Agent, a self-improving coding agent"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/prime-intellect-launches-prime-agent-a-self-improving-coding-agent-siafihers#faq","mainEntity":[{"@type":"Question","name":"What score did Prime Agent get on ARC-AGI-3 and how does it compare to the human expert baseline?","acceptedAnswer":{"@type":"Answer","text":"Prime Agent scored 95.5% on ARC-AGI-3, just above the reported human expert baseline of 95.4%. The run used Claude Opus 5 and relied on code execution, persistent memory, context management, and recursive sub-agents. Whether this complies with the benchmark's official evaluation rules is an open question that some in the community have flagged. Teams benchmarking AI coding agents against ARC-AGI-3 track methodology debates like this on daily.dev."}},{"@type":"Question","name":"What is the Recursive Language Model (RLM) concept used in Prime Agent?","acceptedAnswer":{"@type":"Answer","text":"The Recursive Language Model (RLM) treats context as a variable rather than a fixed window, and uses sub-agent delegation as function calls inside a persistent IPython REPL. This allows the agent to spin up persistent sub-agents, pass messages between them, and track long-running goals with heartbeats and an autonomous eval mode. Developers building multi-agent systems find architectural patterns like RLM discussed on daily.dev."}}]} ```