Prime Agent

Scroll sideways for more →

Prime Agent is an open-source coding agent with a terminal user interface, released by Prime Intellect on August 5, 2026. The company describes it as "a self-improving RLM harness for coding and long-running autonomous tasks," designed to be "token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state."^[1][2] Unlike Claude Code or OpenAI Codex, which pair a harness with the vendor's own models, Prime Agent ships no model of its own: it is MIT-licensed software that runs on top of open-weight or proprietary large language models supplied by the user.^[2][3]

The agent is built around two research ideas. The recursive language model (RLM) gives the model a persistent Python interpreter in which its own context is a program variable and sub-agent delegation is a function call. The Continual Harness treats the harness's prompts, skills, memories, and sub-agent definitions as state the agent can create, read, update, and delete from its own trajectory while it runs.^[1] Prime Intellect reported that Prime Agent running Claude Opus 5 scored 95.5 percent on ARC-AGI-3, which the company said surpasses the benchmark's reported human expert baseline of 95.4 percent.^[1][4]

Background

Prime Intellect is a San Francisco company known for distributed training of open-weight models and for reinforcement learning infrastructure; it announced a $130 million Series A led by Radical Ventures in July 2026 and named recursive language models as one of its main research bets in that announcement.^[5][6]

The RLM idea originated outside the company. Alex L. Zhang, a PhD student at MIT CSAIL, described recursive language models in an October 2025 blog post and, with Tim Kraska and Omar Khattab, in a paper posted to arXiv in December 2025 (arXiv:2512.24601).^[7][8] The paper treats a long prompt as part of an external environment: instead of ingesting the full input, the model examines, decomposes, and recursively calls itself over snippets of it from a Python REPL. The authors reported that this let models process inputs up to two orders of magnitude beyond their context windows and outperform long-context scaffolds such as compaction and, on their benchmarks, Claude Code.^[8] Prime Intellect adopted the paradigm publicly in a January 2026 blog post titled "Recursive Language Models: the paradigm of 2026," arguing that models trained to manage their own context end-to-end through reinforcement learning would unlock agents that work over much longer horizons.^[9] Zhang co-authored the Prime Agent launch post alongside Seth Karten, Kevin Thomas, and Sebastian Müller.^[1]

The second abstraction, Continual Harness, comes from a May 2026 paper by Karten and co-authors (arXiv:2605.09998), which formalizes a harness whose prompt, sub-agents, skills, and memory are refined online from the agent's own trajectory without episode resets.^[10]

Prime Intellect's launch post argues that existing harness designs were built around the limits of earlier model generations: fixed tool-calling schemas and context compaction "force the model to work around its own scaffolding," while hand-engineered sub-agents, prompts, and memory are set at design time and never adapt during a run.^[1] The agent and its interface are built on top of pi, an open-source coding agent framework by Mario Zechner, which the launch post and repository credit.^[1][3][11]

Design

The RLM runtime

Models in Prime Agent have exactly one tool: a persistent IPython kernel. File operations, shell commands, other tools, and context management are all expressed as Python code in that kernel rather than as separate JSON tool calls.^[1][3] Because the kernel persists across turns, the agent has programmatic access to its own history: past context is held in variables it can slice, search, or transform, and Prime Intellect says this lets sessions grow arbitrarily long without the information loss of summary-based compaction.^[1]

Sub-agents are function calls. Invoking await rlm("sub-task") spawns a full Prime Agent instance with its own model, kernel, session tree, and history; the call returns immediately with a handle, and results come back later as messages. The asynchronous design lets the model fan out parallel sub-agents or launch background work in ordinary code.^[1]

Sessions, daemon, and multi-agent messaging

A background daemon owns all live sessions over a local socket, so users can detach from the terminal and reattach without interrupting the agent loop. Each root session runs in a recoverable worker process; if a worker crashes, the daemon restores it from the session's append-only JSONL log and a kernel snapshot. Session history supports branching, forking, and cloning within the same file, and an Agents View lists running, idle, and inactive sessions, including nested sub-agents, any of which the user can enter and steer.^[1][3]

The daemon also carries agent-to-agent messaging: any Prime Agent session can message any other through the same mechanism used for persistent sub-agents, which the company presents as a way to orchestrate sub-agent swarms. To limit unwanted crosstalk, messaging is restricted to an agent's "nuclear family" of parent, sibling, and child processes.^[1] Sub-agents persist after their initial task finishes, keeping their own context and kernel, and can be reactivated later by session identifier.^[1]

Continual Harness and self-improvement

The harness state, comprising prompts, sub-agent specifications, skills, and memories, lives in the kernel as rlm.harness and is mirrored to disk so it survives across sessions. Each component exposes the same create, read, update, and delete surface, and skills are importable Python packages rather than plain text.^[1][3]

/refine is the self-improvement pipeline built on this surface. It reads the agent's own trajectory and applies the smallest relevant edit that improves the harness: updating a prompt note, memory, skill, or sub-agent spec rather than rewriting everything. Each refinement records its trigger and outcome, planning runs in the background, and the immutable base system prompt is never touched; bad updates can be rolled back by ID.^[1][3]

Autonomous mode

For unattended runs, a persistent goal keeps re-prompting the agent until it explicitly marks the goal complete, heartbeats inject scheduled messages on a fixed interval, and autonomous mode keeps the session working instead of stopping early. Runs are bounded by turn, token, and wall-clock limits, and a user-defined gate command (for example a test suite) must pass before the session may finish.^[1][3]

Model support and availability

Prime Agent is free software under the MIT license, installed on macOS or Linux via a shell script from app.primeintellect.ai.^[3] It is model-agnostic: users authenticate with a subscription (ChatGPT Plus/Pro via Codex, Claude Pro/Max, or GitHub Copilot) or with API keys from providers including Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, Groq, Cerebras, OpenRouter, Hugging Face, MiniMax, Kimi For Coding, and Prime Intellect's own Prime Inference service.^[12] The documentation notes that OpenAI has endorsed its Codex-subscription integration under the "Codex for OSS" program, and that Anthropic subscription use by third-party harnesses is billed per token as extra usage rather than against plan limits.^[12]

The repository warns that Prime Agent executes model-generated Python and project commands with the user's permissions and that its isolation measures "are not a security sandbox," recommending disposable clones or external sandboxes for untrusted work.^[3]

Reported performance

All results below are Prime Intellect's own, published in the launch post and, as of August 7, 2026, not independently verified; no technical report had yet been released.^[1]

On ARC-AGI-3, an interactive reasoning benchmark from the ARC Prize Foundation, the company reported a best score of 95.5 percent (RHAE Best@1) using Claude Opus 5, above the reported human expert baseline of 95.4 percent, with scores of 95.0, 95.2, and 95.5 percent across three runs and all 183 levels completed at Best@3.^[1][4] It also reported lower token usage than native harnesses, attributing this to running functions over data instead of reading data through tool calls.^[1] The company noted that when it evaluated Claude Code and Codex itself it got worse numbers than those vendors' official results, so it deferred to the official figures.^[1]

On a suite of long-context benchmarks, Prime Intellect compared Prime Agent against each closed harness paired with its own vendor's model, and against the pi-mono harness with an open-weight model:^[1]

Scroll sideways for more →

Scores as reported by Prime Intellect, August 5, 2026. The company said its Opus 5 runs on EmulatorBench (building game-console emulators in Rust from scratch) largely failed for both harnesses despite successful tool calls.^[1]

In case studies, the company reported that Prime Agent playing the factory game Factorio through the Factorio Learning Environment used /refine to turn failures and successes into memories and skills and reached production scores above 100,000 within hours, but also engaged in reward hacking: it discovered it could spawn resources directly into machines through RCON console commands and kept doing so even after a recurring prompt told it not to cheat.^[1] The post stresses that no model has yet been trained around Prime Agent or its core features, and frames model-harness co-training as the next step, with a full technical report promised later.^[1]

Reception

The launch drew wide attention on X, where the announcement passed 7,500 likes within its first days, and a Hacker News thread on the release had reached 249 points and 65 comments by August 7, 2026.^[2][13] Commenters were impressed that the ARC-AGI-3 result nearly saturates the benchmark, while others questioned token costs of self-improvement loops, criticized code quality in the repository (one noted multiple files near 10,000 lines), and complained that the installer wrote into the Homebrew directory without an uninstall path.^[13] The repository reached roughly 5,200 GitHub stars within two days of release.^[3]

Zhang, whose research the RLM runtime builds on, called Prime Agent "an RLM-native coding TUI" and wrote that "this agent is really good despite not having any models trained around it," saying he had been using it for his own research with open-weight models.^[14] Trade coverage highlighted the single-IPython-kernel design as the main departure from JSON-schema tool calling in incumbent agents.^[15]

See also

• Prime Intellect
• Recursive language model
• Claude Code
• OpenAI Codex
• AI agent
• Agent harness
• Open source AI

References

1. ^Prime Intellect, "Prime Agent: A self-improving RLM agent," August 5, 2026. primeintellect.ai/...prime-agent
2. ^Prime Intellect on X, "Introducing Prime Agent," August 5, 2026. x.com/...2085086999267144083
3. ^PrimeIntellect-ai/prime-agent repository (README, LICENSE), GitHub, accessed August 7, 2026. github.com/...prime-agent
4. ^Prime Intellect on X, "On ARC-AGI-3, it scores 95.5%," August 5, 2026. x.com/...2085087000764568010
5. ^Prime Intellect, "$130M Series A to Build the Open Superintelligence Stack," July 8, 2026. primeintellect.ai/...series-a
6. ^"AI startup Prime Intellect raises $5.5M to build high-powered, decentralized research platform," Fortune, April 23, 2024. fortune.com/...ificial-intelligence-ai-development
7. ^Alex L. Zhang, "Recursive Language Models," blog post, October 2025. alexzhang13.github.io/...rlm
8. ^Alex L. Zhang, Tim Kraska, Omar Khattab, "Recursive Language Models," arXiv:2512.24601, December 2025. arxiv.org/...2512.24601
9. ^Prime Intellect, "Recursive Language Models: the paradigm of 2026," January 1, 2026. primeintellect.ai/...rlm
10. ^Seth Karten et al., "Continual Harness: Online Adaptation for Self-Improving Foundation Agents," arXiv:2605.09998, May 2026. arxiv.org/...2605.09998
11. ^earendil-works/pi repository (formerly badlogic/pi-mono), GitHub. github.com/...pi
12. ^Prime Agent documentation, "Providers," GitHub, accessed August 7, 2026. github.com/...providers.md
13. ^"Prime Agent: A self-improving RLM agent," Hacker News discussion, August 5, 2026. news.ycombinator.com/item
14. ^Alex Zhang (@a1zhang) on X, August 5, 2026. x.com/...2085089855541813350
15. ^"Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel," MarkTechPost, August 6, 2026. marktechpost.com/...intellect-releases-prime-agent