Prime Agent: an open-source coding agent that rewrites its own harness
Open-source, one curl to install. It edges the human ARC-AGI-3 baseline, runs persistent sub-agents, and refines itself. Plus Zed DeltaDB and Neon Castform.
Copy markdown
One line to install, and it's free
Prime Intellect open-sourced Prime Agent on Aug 5: one curl command drops a full coding agent onto your machine, no API lock-in. It drives tools programmatically from a persistent IPython REPL instead of a fixed tool menu.
It edits its own prompts, skills, and memory
The 'Continual Harness' lets the agent rewrite its own operational state mid-run. A /refine pipeline updates prompts, skills, memory, and sub-agents from the failures it just hit — self-improvement you can read, not a black box.
Sub-agents that survive across sessions
Its Recursive Language Model treats context as a variable and spawns sub-agents as function calls. They persist across sessions and message each other, with goal-setting, heartbeats, and token budgets for long autonomous runs.
95.5% on ARC-AGI-3, past the human baseline
Running on Claude Opus 5, Prime Agent reports 95.5% RHAE Best@1 on ARC-AGI-3 — just past the 95.4% human-expert mark. Take one lab's own number with salt, but the whole harness is open, so you can reproduce it yourself.
Zed's DeltaDB gives every edit a stable identity
Zed opened early access to DeltaDB: version control that records every operation between commits and links each change to the agent conversation that made it, plus live collaborative branching before anyone commits. Waitlist only for now.
Post-train an open model on your own data
Neon and Castform showed a 4B open model RL-tuned to match GPT-5.6 Sol on retrieval at ~100x lower cost, running the whole loop on Neon's Lakebase Postgres. It's a platform pitch, not open weights — but 'turn your corpus into RL tasks' is the takeaway.