Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel

Prime Intellect has open-sourced Prime Agent, a self-improving coding harness designed around two abstractions, the Recursive Language Model (RLM) and Continual Harness. Fixed tool schemas and context compaction force a model to work around its own scaffolding. Prime Agent replaces both with a persistent Python REPL and a rewritable harness. With Opus 5, it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%. It is MIT-licensed.

Is it deployable

Yes, today. Prime Agent installs on Linux or macOS with one command. It runs on subscription logins (Codex, Claude Pro/Max, GitHub Copilot), API keys (Anthropic, OpenAI, Google, Groq, Fireworks, Prime Inference, and others), Azure OpenAI, Amazon Bedrock, and self-hosted vLLM, Ollama, or LM Studio endpoints. Self-hosting an open-weights model such as GLM-5.2 keeps code inside your own network.

Company level: Best fit is mid-size to large engineering orgs and AI labs that already run isolated CI containers. Prime Intellect states plainly that worker and kernel processes are not a security sandbox. Deployment therefore needs disposable clones or restricted environments. Solo developers can install it, but the payoff appears on multi-hour tasks.
Industries: Developer tooling, semiconductor and HPC teams writing GPU kernels, simulation and gaming, quantitative research, and AI research labs.
Applications: Overnight refactors behind a test gate, spec-driven builds from scratch, kernel optimization, long-horizon agent evaluation, and autoresearch.

What Prime Intellect shipped

Prime Agent is built on two abstractions. The Recursive Language Model (RLM) treats context as a variable and sub-agent delegation as function calls inside a REPL. The Continual Harness treats prompts, sub-agents, skills, and memory as state the agent can create, read, update, and delete from its own trajectory. Both papers have Prime Agent authors on them. The TUI is built on pi.

= CODE.length){ repl.innerHTML += ‘‘; return; } repl.innerHTML += CODE[i] + ‘ ‘; if(i === 1){ setNode(‘root’,’p-run’,’running’,’turn in progress’); setNode(‘auth’,’p-run’,’running’,’admitted’); } if(i === 2){ setNode(‘api’,’p-run’,’running’,’admitted’); } i++; ping(); later(line, reduce ? 60 : 620); } line(); later(function(){ setNode(‘auth’,’p-done’,’replied’,’agent_message → parent’); tHint.textContent=”auth-expert replied. The parent never blocked.”; ping(); }, reduce ? 400 : 2600); later(function(){ setNode(‘api’,’p-done’,’replied’,’agent_message → parent’); }, reduce ? 500 : 3600); later(function(){ setNode(‘root’,’p-idle’,’idle’,’2 replies merged’); tHint.textContent=”Both children retained on disk. Message them again any time.”; ping(); }, reduce ? 600 : 4300); } document.getElementById(‘runTurn’).addEventListener(‘click’, runTurn); document.getElementById(‘resetTurn’).addEventListener(‘click’, resetTurn); document.querySelectorAll(‘.node’).forEach(function(n){ n.addEventListener(‘click’, function(){ document.querySelectorAll(‘.node’).forEach(function(x){ x.classList.remove(‘sel’); }); n.classList.add(‘sel’); }); }); /* ———- pane 2: harness ———- */ var counts = {p:0,g:0,k:0,m:0}, history = []; var WRITES = { p: [‘create_prompt_note(“check the workspace diff before every gate rerun”)’], g: [‘create_subagent(“auth-reviewer”, spec=”read-only review of auth/ changes”)’], k: [‘create_skill(“retry helper”, reference={“type”:”python”,”import”:”retry_helper”})’], m: [‘create_memory(“flaky test pattern”, “retry three times before failing”)’] }; var NAMES = {p:’\u03C1 prompt’, g:’G sub-agents’, k:’K skills’, m:’M memory’}; var hlog = document.getElementById(‘hlog’); function logLine(html){ var d = document.createElement(‘div’); d.innerHTML = html; hlog.appendChild(d); while(hlog.children.length > 6) hlog.removeChild(hlog.firstChild); ping(); } function bump(kind, text, viaRefine){ counts[kind]++; document.querySelector(‘[data-c=”‘+kind+'”]’).textContent = counts[kind] + (counts[kind]===1?’ entry’:’ entries’); var card = document.querySelector(‘.hcard[data-h=”‘+kind+'”]’); card.classList.remove(‘hit’); void card.offsetWidth; card.classList.add(‘hit’); history.push(kind); logLine(‘‘+(viaRefine?’/refine’:’rlm.harness’)+’ ‘+text+’ ✓ written to ‘+NAMES[kind]+’‘); } document.querySelectorAll(‘.hcard’).forEach(function(c){ c.addEventListener(‘click’, function(){ var k = c.getAttribute(‘data-h’); bump(k, WRITES[k][0], false); }); }); var REFINES = [ {k:’m’, t:’trigger: same test failed twice → smallest edit: add memory’}, {k:’k’, t:’trigger: retry pattern reused 3× → promote to skill’}, {k:’p’, t:’trigger: gate rerun on unchanged workspace → add prompt note’}, {k:’g’, t:’trigger: repeated review sub-task → save sub-agent spec’} ]; var ri = 0; document.getElementById(‘refineBtn’).addEventListener(‘click’, function(){ var r = REFINES[ri % REFINES.length]; ri++; logLine(‘/refine planning in background … conversation not blocked’); later(function(){ bump(r.k, r.t, true); }, reduce ? 80 : 850); }); document.getElementById(‘rollbackBtn’).addEventListener(‘click’, function(){ var k = history.pop(); if(!k){ logLine(‘rollback nothing to revert’); return; } counts[k]–; document.querySelector(‘[data-c=”‘+k+'”]’).textContent = counts[k] + (counts[k]===1?’ entry’:’ entries’); logLine(‘rollback reverted last edit to ‘+NAMES[k]+’ by id — base system prompt untouched’); }); /* ———- pane 3: evals ———- */ var drawn = false; function drawBars(){ if(drawn) return; drawn = true; document.querySelectorAll(‘#bars .fill’).forEach(function(f, i){ later(function(){ f.style.width = f.getAttribute(‘data-w’) + ‘%’; }, reduce ? 0 : 120 * i); }); } var DATA = { glm: {rival:’Pi-mono (w/ sub-agents)’, rows:[ [‘OOLONG (yahoo, 128k)’, 0.700, 0.420], [‘OOLONG-Pairs’, 0.874, 0.556], [‘OBLIQ-Bench (math)’, 0.669, 0.635], [‘LongBenchPro (English)’, 0.777, 0.768], [‘LongBenchv2’, 0.680, 0.696], [‘ManyIH Coding’, 0.424, 0.386], [‘ManyIH IF’, 0.209, 0.164], [‘LongCot-Mini’, 0.638, 0.613], [‘EmulatorBench’, 0.208, 0.000] ]}, opus: {rival:’Claude Code’, rows:[ [‘OOLONG (yahoo, 128k)’, 0.900, 0.920], [‘OOLONG-Pairs’, 0.929, 0.922], [‘OBLIQ-Bench (math)’, 0.802, 0.795], [‘LongBenchPro (English)’, 0.804, 0.790], [‘LongBenchv2’, 0.744, 0.746], [‘ManyIH Coding’, 0.536, 0.522], [‘ManyIH IF’, 0.225, 0.175], [‘LongCot-Mini’, 0.722, 0.558], [‘EmulatorBench*’, 0.047, 0.062] ]}, gpt: {rival:’Codex’, rows:[ [‘OOLONG (yahoo, 128k)’, 0.940, 0.500], [‘OOLONG-Pairs’, 0.911, 0.895], [‘OBLIQ-Bench (math)’, 0.612, 0.646], [‘LongBenchPro (English)’, 0.794, 0.790], [‘LongBenchv2’, 0.714, 0.704], [‘ManyIH Coding’, 0.499, 0.454], [‘ManyIH IF’, 0.216, 0.232], [‘LongCot-Mini’, 0.671, 0.681], [‘EmulatorBench’, 0.275, 0.228] ]} }; function fmt(v){ return v.toFixed(3); } function renderTable(key){ var d = DATA[key], tb = document.getElementById(‘tbody’); document.getElementById(‘rivalHead’).textContent = d.rival; tb.innerHTML = d.rows.map(function(r){ var aWin = r[1] > r[2], bWin = r[2] > r[1]; return ‘<tr><td>‘+r[0]+’</td>‘+ ‘<td class=""'+(aWin?'win-v':'lose-v')+'"">‘+fmt(r[1])+’</td>‘+ ‘<td class=""'+(bWin?'win-v':'lose-v')+'"">‘+fmt(r[2])+’</td></tr>‘; }).join(”); ping(); } document.querySelectorAll(‘.seg button’).forEach(function(b){ b.addEventListener(‘click’, function(){ document.querySelectorAll(‘.seg button’).forEach(function(x){ x.setAttribute(‘aria-pressed’,’false’); }); b.setAttribute(‘aria-pressed’,’true’); renderTable(b.getAttribute(‘data-m’)); }); }); renderTable(‘glm’); /* ———- pane 4: autonomous ———- */ var steps = [].slice.call(document.querySelectorAll(‘.step’)); var turns = 0; function clearSteps(){ steps.forEach(function(s){ s.className=”step”; }); } function setBudget(t){ turns = t; document.getElementById(‘bvTurns’).textContent = t + ‘ / 20’; document.getElementById(‘bpTurns’).style.width = (t/20*100) + ‘%’; document.getElementById(‘bpTok’).style.width = Math.min(t*9, 100) + ‘%’; document.getElementById(‘bpTime’).style.width = Math.min(t*7, 100) + ‘%’; } function runLoop(){ clearTimers(); clearSteps(); setBudget(0); var seq = [ [0,’on’,0], [1,’on’,1], [2,’on’,2], [3,’fail’,3], [1,’on’,4], [2,’on’,5], [4,’pass’,6] ]; seq.forEach(function(s, i){ later(function(){ clearSteps(); steps[s[0]].className=”step ” + s[1]; if(s[0] === 1) setBudget(turns + 1); if(s[1] === ‘fail’) document.getElementById(‘bvTok’).textContent=”gate output returned”; if(s[1] === ‘pass’){ document.getElementById(‘bvTok’).textContent=”within budget”; } }, reduce ? i*60 : i*900); }); } document.getElementById(‘runLoop’).addEventListener(‘click’, runLoop); document.getElementById(‘stopLoop’).addEventListener(‘click’, function(){ clearTimers(); clearSteps(); setBudget(0); document.getElementById(‘bvTok’).textContent=”bounded”; }); ping(); })();

Source link