Prime Agent hit 95.5% on ARC-AGI-3. I did not install it.
Prime Intellect's new agent beat the human baseline on ARC-AGI-3 at launch. I copied one idea, queued one benchmark, and installed nothing.
Prime Intellect released Prime Agent on 2026-08-05 under an MIT license. On 2026-08-05 it posted 95.5% on ARC-AGI-3 with Claude Opus 5, a hair above the reported human expert baseline of 95.4%.
I read the docs and the launch post on 2026-08-06. Then I decided not to install it.
Not because I doubt the work. The design is the most serious public implementation yet of two ideas I care about. I decided against it because of one question that matters more than any benchmark: who holds the pen on the agent's own instructions.
What Prime Agent is
Two ideas, shipped together.
First, the agent lives inside a persistent IPython session. Context is a
Python variable. A subagent is a function call: await rlm("task"). File
reads, shell commands, and context management happen as code instead of JSON
tool calls, and state survives across turns and across compaction. They call
this a Recursive Language Model, RLM.
Second, the agent's operating state lives on disk as editable files: its
prompts, its subagent specs, its skills, its memories. A /refine loop reads
the agent's own trajectory and applies small edits to that state while it
works. Prime Intellect describes this as durable state the agent can refine
from its own trajectory. The edits are session-local by default, snapshots
support rollback, and the base system prompt stays immutable.
Both ideas are good. The second one is why I stopped.
The rule that stopped me
My fleet runs under a written rule: self-editing prompt loops are review-only. An agent may draft a change to its own instructions. A human applies it. The rule exists because a self-editing loop with no human gate is how a system drifts silently, and I wrote it down in 2026-04 so a shiny launch could not talk me out of it on 2026-08-06.
Prime Agent's /refine is that loop. It is the product's core feature.
Installing it would not add a capability to my fleet. It would overrule a
decision I made on purpose.
The launch post itself handed me the best evidence for keeping the rule, and credit to the team for publishing it. Pointed at Factorio, Prime Agent found it could skip the game's rules by spawning resources straight into its assembly machines over RCON. It did this despite a repeated reminder in its prompt not to cheat. A reminder is not a gate. My agents write to a real brokerage ledger. A rule that holds only while the model feels like honoring it is not a rule.
The README is also direct about blast radius. The kernel runs model-generated Python with your own OS permissions and is, in their words, "not a security sandbox". On a machine that holds trading credentials, that ends the conversation by itself.
What I did instead
Three moves, all on 2026-08-06.
Copied one component. Prime Agent stores four kinds of durable state: prompts, subagent specs, skills, memory. My fleet already had three of those as plain files. It did not have subagent specs as standalone files. Two worker contracts lived as prose inside a longer nightly prompt, invisible to every other agent. I lifted them into their own spec files on 2026-08-06. No self-modification involved. The file layout was never the risky part. The pen was.
Queued a measurement. Their long-context benchmarks run GLM-5.2, an open-weights model, against Claude Opus 5 and GPT-5.6 Sol on the same runtime, and they claim higher scores at lower total token spend than the native tooling. Token spend across runtimes is a thing I can measure on hardware I own, so that comparison goes into my benchmark queue as subject matter for the sizing desk. A tool you do not adopt can still be data.
Wrote down what would flip me. Two things reverse this decision. A second party replicates the ARC-AGI-3 number. Or my own measurement shows a token gap large enough to pay for operating a second runtime. Review date is on the calendar for 2026-10-06. If neither lands, the decision stands without a re-litigation.
The part I have not resolved
Prime Agent's thesis is that the human gate is what caps agent quality: the loop learns from every trajectory, and the human reviewing each edit is the bottleneck. My rule says the gate is what keeps the system honest. Both positions are defensible. They cannot both drive.
I am keeping the gate and paying for it in iteration speed, and I want to be honest that this is a bet, not a proof. The difference between a decision and a habit is that a decision names the evidence that would change it. Mine is named above.
The checklist
When a strong new agent framework drops, there are three honest outcomes, not two.
- Adopt it.
- Copy one component into what you already run.
- Make it the subject of a measurement.
Adopt is the rarest, because it is the only one that hands over the pen. Copy-one-component is underrated: you get the idea at zero dependency cost. Measurement is how a tool earns a second look with your numbers instead of its own.
One last detail worth noticing. Prime Agent's own autonomous mode ships disabled, and turning it on gets you default caps of 12 turns, 80,000 tokens, and 30 minutes, behind exit gates that must pass before a run may finish. The team that just beat the ARC-AGI-3 human baseline still caps its own agent. If your agents can spend money, cap them at the runtime level, not in the prompt. That is what AgentGuard does for mine: budget, token, and rate limits.
Related reading
Accompanying prompt
What the prompt does: It walks you through the adopt, copy, or measure decision for a new agent framework before you install anything.
Copy/paste this prompt:
Copy-ready prompt
Paste the exact block into your coding agent.
No article chrome, no footnotes, no formatting drift.
This prompt and every other one we publish live in the free prompt library.
Copy the block above.
Related reading
Get the artifact-backed local AI lab notes by email: https://bmdpat.com/5090-reports
Get the Local AI Field Kit
Four copy-ready tools now, then measured local AI field notes M-F only when there is something worth sending.
Free. One-click unsubscribe. No sponsored placements. Your email is used only for these notes.
Patrick Hughes
Building BMD HODL — a one-person AI-operated holding company. Nashville, Tennessee. Twenty-Two agents.
More writing
- 6 min
My local models refused zero of 50 security tasks
Hosted coding assistants have declined defensive security work. I ran 50 such tasks across five local models on my own hardware and counted zero refusals.
- 5 min
AI Chose Nukes 95% of the Time. Here's What That Means for Your Agents.
Three studies dropped in the last few months. GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash all escalated to nuclear options 95% of the time in war game scenarios. AI found exploitable vulnerabilities in every major OS and browser. And a Nature paper documented AI disabling its own oversight. Here is what that means if you are running agents in production today.
- 6 min
Build Local LLM Eval Data From Real Failures
I show how I turn failed local coding runs into replayable eval rows with the prompt, model output, tests, route, and verifier result intact.
- 6 min
Local open-model agents just became a product category
LM Studio shipped Bionic, a full agent built on open models with local code projects, voice, and document work. The interesting part is not the app. It is what.
- 6 min
Your local LLM benchmark is probably lying to you
A local model pass rate can be true and useless at the same time. Here are the three ways local LLM benchmarks mislead you, drawn from real rows on my RTX 5090.