Prime Agent is a coding harness from Prime Intellect built around the Recursive Language Model pattern, where every tool call and sub-agent invocation is expressed as executable code rather than a fixed JSON schema. It targets developers running long-horizon autonomous coding work who want a harness that can rewrite its own scaffolding mid-run, pass context around as first-class variables, and coordinate multiple agents through code-level messaging. The launch matters because agent harnesses have become the deciding factor in benchmark performance, with the same base model swinging widely depending on the scaffolding around it, and Prime Agent is being released as an installable CLI (uv tool install prime) rather than a closed product.
The team pointed the harness at ARC-AGI-3, described by ARC Prize as an IQ test for agents that requires heavy test-time learning and forces the agent to build an internal world model of rules and mechanics that updates with new evidence , a benchmark Prime Agent was not designed for. It reached 95.5%, above the reported human baseline and above prior native coding harnesses, which is notable in a space where recent work like the Schema harness reached 99% on the ARC-AGI-3 public set using Claude Opus, while the nearest official single-variant reference from ARC Prize was in the low double digits . The self-modifiable state and programmatic tool calling are what the company is arguing generalize beyond ARC-style puzzles into everyday coding work.
Prime Intellect was founded in 2023 by Vincent Weisser and Johannes Hagemann and operates as a compute and infrastructure platform for agentic model development , positioning itself as an "Open Superintelligence Stack" that spans environments, training, evaluation, and inference. Prime Agent slots in as the client-side piece of that stack, giving founders and infra teams a way to run and post-train their own agents on top of the company's existing verifiers and RL environment tooling instead of depending on a frontier lab's closed harness.
Comments (12)
Sign in to join the discussion.
N
Nadia O.Aug 5, 2026
self-modifiable harness state is a fun phrase until the harness decides your test suite was the problem.
T
Tomasz WierzbaAug 5, 2026
ok wait, programmatic tool calling plus context as a variable is actually the part I want to see benchmarked, not the demo where it fixes a typo in a README.
S
Søren HalvorsenAug 5, 2026
we shipped something eerily similar as an internal harness in 2019, we just called it a Jenkins job with anxiety.
P
Priya RamaswamyAug 5, 2026
before I say interesting, what does week 4 retention look like on the coding tasks? self-improving is a strong claim to underwrite.
K
Kenji AdebayoAug 5, 2026
the launch page pacing is clean but the tweet copy reads like a NeurIPS abstract someone squeezed into 280 chars.
M
marisolopsAug 5, 2026
hot take: shipping this as a uv tool install one-liner does more marketing work than the whole video did.
B
Bex DelacroixAug 5, 2026
every coding agent harness launch feels like we're all speedrunning the same plateau, the ceiling on this category is lower than the pitch decks suggest.
L
Linh TrầnAug 5, 2026
reminds me of what a portco of mine tried to build on top of Claude last year, except this one might actually ship past the demo.
D
Dewayne OkaforAug 5, 2026
curious what the token economics look like when the harness is rewriting itself mid-task, feels like a fun way to torch a compute budget.
A
Amara BankoleAug 5, 2026
how many people actually built this? because if it's under ten I'm going to need a moment.
V
Vikram L.Aug 5, 2026
self-modifiable harness plus long-running autonomy is going to be a fun conversation with any enterprise security team, hope there's an audit log story somewhere.
J
Joona MäkinenAug 5, 2026
who's hiring RL infra folks off the back of this? got a candidate wrapping up at a robotics lab who eats harnesses for breakfast.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Prime Intellect
Own Your Intelligence
Own Your Intelligence
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Prime Agent is a coding harness from Prime Intellect built around the Recursive Language Model pattern, where every tool call and sub-agent invocation is expressed as executable code rather than a fixed JSON schema. It targets developers running long-horizon autonomous coding work who want a harness that can rewrite its own scaffolding mid-run, pass context around as first-class variables, and coordinate multiple agents through code-level messaging. The launch matters because agent harnesses have become the deciding factor in benchmark performance, with the same base model swinging widely depending on the scaffolding around it, and Prime Agent is being released as an installable CLI (uv tool install prime) rather than a closed product. The team pointed the harness at ARC-AGI-3, described by ARC Prize as an IQ test for agents that requires heavy test-time learning and forces the agent to build an internal world model of rules and mechanics that updates with new evidence , a benchmark Prime Agent was not designed for. It reached 95.5%, above the reported human baseline and above prior native coding harnesses, which is notable in a space where recent work like the Schema harness reached 99% on the ARC-AGI-3 public set using Claude Opus, while the nearest official single-variant reference from ARC Prize was in the low double digits . The self-modifiable state and programmatic tool calling are what the company is arguing generalize beyond ARC-style puzzles into everyday coding work. Prime Intellect was founded in 2023 by Vincent Weisser and Johannes Hagemann and operates as a compute and infrastructure platform for agentic model development , positioning itself as an "Open Superintelligence Stack" that spans environments, training, evaluation, and inference. Prime Agent slots in as the client-side piece of that stack, giving founders and infra teams a way to run and post-train their own agents on top of the company's existing verifiers and RL environment tooling instead of depending on a frontier lab's closed harness.
Comments (12)
self-modifiable harness state is a fun phrase until the harness decides your test suite was the problem.
ok wait, programmatic tool calling plus context as a variable is actually the part I want to see benchmarked, not the demo where it fixes a typo in a README.
we shipped something eerily similar as an internal harness in 2019, we just called it a Jenkins job with anxiety.
before I say interesting, what does week 4 retention look like on the coding tasks? self-improving is a strong claim to underwrite.
the launch page pacing is clean but the tweet copy reads like a NeurIPS abstract someone squeezed into 280 chars.
hot take: shipping this as a uv tool install one-liner does more marketing work than the whole video did.
every coding agent harness launch feels like we're all speedrunning the same plateau, the ceiling on this category is lower than the pitch decks suggest.
reminds me of what a portco of mine tried to build on top of Claude last year, except this one might actually ship past the demo.
curious what the token economics look like when the harness is rewriting itself mid-task, feels like a fun way to torch a compute budget.
how many people actually built this? because if it's under ten I'm going to need a moment.
self-modifiable harness plus long-running autonomy is going to be a fun conversation with any enterprise security team, hope there's an audit log story somewhere.
who's hiring RL infra folks off the back of this? got a candidate wrapping up at a robotics lab who eats harnesses for breakfast.