StartPrime Agent Is Free, Open Source, and Remembers What Every Other Coding Agent Forgets
By Addy · August 8, 2026 · Editorial standards
Prime Intellect open-sourced Prime Agent on August 5, three days after closing a $130 million Series A that pushed the company past a $1 billion valuation. The launch number everyone is repeating is 95.5% on ARC-AGI-3, run with Claude Opus 5, edging past the reported human expert baseline of 95.4%. That number is worth a skeptical look on its own, and it comes later. The more useful way to understand Prime Agent is to first look at what every other coding agent on the market does, and then look at the one thing Prime Agent does instead.
How Every Other Coding Agent Works
Claude Code, Codex, Cursor, and Cline all operate on the same basic shape: a fixed tool-calling schema, a context window managed through compaction as it fills up, and a session that ends when you close it. That last part is the practical daily annoyance. You open a new session, the agent has no memory of your conventions, your test-first preference, or the pattern you corrected it on yesterday, and you explain it again. Every one of these tools is, underneath its polish, stateless by design. The context window is the entire memory, and the memory resets.
What Prime Agent Does Differently
Prime Agent replaces that fixed schema with something closer to a workspace than a window. Its only tool is a persistent IPython kernel, and instead of reading data into a limited context to reason over it, the model writes and runs code directly against it, treating context as something it can program over rather than a container it has to fit inside. Sub-agents aren't a separate feature bolted on top, they're function calls, await rlm("subtask"), that spin up a full independent session, run in the background, and report back through a message or a file, surviving even if the parent session disconnects.
The more consequential piece is what Prime Intellect calls the Continual Harness. A /refine command lets the agent rewrite its own system prompt, save patterns it notices to memory, and adjust its own sub-agent specifications, mid-task, based on what is and isn't working, with every change logged and reversible by ID. Picture the difference between someone re-reading and re-summarizing an entire case file every time they need to make one small decision, versus someone with an actual notebook they keep updating, one they can hand specific pages to an assistant and get a report back on. That second version is closer to what Prime Agent is attempting. Whether the current generation of models actually benefits from having that capability is a separate question. Prime Intellect's own writeup is candid about this: many of the harness's features aren't fully used by a model that wasn't trained alongside it, and the company frames real gains as still ahead, once models and harness are trained together rather than paired after the fact.
The Benchmark Claim, and What It Doesn't Cover
The 95.5% ARC-AGI-3 result is real in the sense that Prime Intellect published it, but it hasn't been independently reproduced, and some in the community have already questioned whether the run follows the benchmark's official evaluation methodology. Prime Intellect's broader claim, that Prime Agent improves on a model's own native harness across models generally, not just on this one benchmark, is the more interesting claim and the harder one to verify from outside, since it launched days ago.
Worth knowing for context: Meta shipped a competing terminal coding agent, Muse Code, the same day, built as a closed, co-trained product paired specifically with its own model rather than an open harness for any model. Nobody has run both agents on the same benchmark yet, so any head-to-head ranking implied by the coincidence of the launch date doesn't actually exist. That's a useful reminder that a lot of what circulates about new agent tools in the first week is timing and framing, not measurement.
Pricing and Setup
Prime Agent itself is free. It's MIT licensed, there's no Prime Intellect subscription required to use it, and the only cost is whatever model access you already have or choose to pay for, a Claude Pro or Max subscription, ChatGPT Plus or Pro through Codex, GitHub Copilot, or an API key from Anthropic, OpenAI, Google, or several smaller providers. It also connects to self-hosted models through Ollama, vLLM, or LM Studio, which costs nothing beyond your own hardware and, for anyone concerned about code leaving their network at all, is the most complete privacy option available here.
Installation is a single command on macOS or Linux, curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh, followed by /login to connect a provider. There's no Windows support at launch. A current Node.js install is required, and building from source needs Node.js 22.8.0 or newer.
The Part Worth Being Careful About
Prime Intellect states this plainly in its own documentation rather than burying it: Prime Agent executes model-generated code and commands with your actual user permissions, and its process isolation is built for recovery and continuity, not security. It is not a sandbox. The project's own guidance is to run it against a disposable clone or a clean, restorable checkpoint, and to avoid pointing it at untrusted repositories, instructions, or extensions without an external sandbox around it. Given how much of this year's AI security news has been about agents escaping containment that was supposed to hold them, a widely adopted tool that skips the sandbox by design and says so upfront deserves to be taken at its word, in both directions, believed about the risk and given credit for disclosing it.
Where That Actually Leaves It
The fairest verdict on Prime Agent right now isn't "best coding agent," it's narrower than that. Claude Code still leads on raw coding output quality for most people's daily work. Cursor wins on IDE integration. Cline wins on simplicity and an approval-first workflow that suits people who want to review every step. Prime Agent's actual advantage is persistence, an agent that keeps what it learned about your project instead of forgetting it every session, and that's a genuinely useful, currently uncommon capability among the major open tools in this category, not a claim that it beats everything else at everything else.
Previously on TheQuery: