[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-74200":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":9,"rankLanguage":9,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":24,"hasPages":22,"topics":25,"createdAt":9,"pushedAt":9,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":15,"starSnapshotCount":15,"syncStatus":14,"lastSyncTime":29,"discoverSource":30},74200,"atlas-gic","chrisworsey55\u002Fatlas-gic","chrisworsey55","ATLAS by General Intelligence Capital — Self-improving AI trading agents using Karpathy-style autoresearch",null,"Python",1940,357,28,2,0,8,36,119,24,20.66,"Other",false,"main",true,[],"2026-06-12 02:03:23","## 🚀 ATLAS Agents — Launching Next Week\n\nThe platform built on top of this repo is launching next week.\n\n**Three ways to use it:**\n\n1. **Copy ATLAS** — trade alongside the autoresearch system \n   in this repo. Live signals, auto-execution, full \n   transparency. Up 30% since launch.\n\n2. **Build your own** — describe a strategy in plain English, \n   backtest on 18 months of real data, deploy live.\n\n3. **Marketplace** — publish your agent, earn when others \n   copy it.\n\nSIMONS is already live on Kalshi prediction markets. \n60% win rate.\n\nAs a thank you for all the feedback, ideas, and support \nthat helped shape this project — we're giving the GitHub \ncommunity 20% off. Forever, not just the first month.\n\n→ **[atlasagents.co](https:\u002F\u002Fatlasagents.co)**\nUse code **GITHUB20** at checkout. First 100 only.\n\n---\n\n# ATLAS - Self-Improving AI Trading Agents\n\n**Built by [General Intelligence Capital](https:\u002F\u002Fgeneralintelligencecapital.com)**\n\r\n[Karpathy's autoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch) + [Soros's reflexivity](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FReflexivity_(social_theory)) + [MiroFish swarm simulation](https:\u002F\u002Fgithub.com\u002F666ghj\u002FMiroFish) applied to financial markets.\r\n\r\nThe agent prompts are the weights. Sharpe ratio is the loss function. No GPU needed.\r\n\r\n⭐ 944 stars · 202 forks\r\n\r\n---\r\n\r\n## What Is This?\r\n\r\nATLAS is a framework for autonomous AI trading agents that improve their own prompts through market feedback, train on different market regimes, create new agents when they detect knowledge gaps, and simulate reflexive futures to prepare for what's coming.\r\n\r\n25+ agents debate markets daily across 4 layers. Every recommendation is scored against real outcomes. The worst-performing agent gets its prompt rewritten. If performance improves, the git commit survives. If not, git revert.\r\n\r\nNow running live with real capital.\r\n\r\n---\r\n\r\n## Architecture\r\n\r\n### Layer 1 - Macro (10 agents)\r\n\r\nCentral bank, geopolitical, China, dollar, yield curve, commodities, volatility, emerging markets, news sentiment, institutional flow.\r\n\r\nThese agents set the regime. Risk on or risk off? What's the macro backdrop?\r\n\r\n### Layer 2 - Sector Desks (7 agents)\r\n\r\nSemiconductor, energy, biotech, consumer, industrials, financials, plus a Bloomberg-style relationship mapper that tracks supply chains, ownership, analyst coverage, and competitive dynamics.\r\n\r\n### Layer 3 - Superinvestors (4 agents)\r\n\r\n* **Druckenmiller** - macro\u002Fmomentum: what's the big asymmetric trade?\r\n* **Aschenbrenner** - AI\u002Fcompute: who benefits from the capex cycle?\r\n* **Baker** - deep tech\u002Fbiotech: who has real IP moats?\r\n* **Ackman** - quality compounder: pricing power + FCF + catalyst?\r\n\r\n### Layer 4 - Decision (4 agents)\r\n\r\n* **CRO** - adversarial risk officer: attacks every idea, finds correlated risks\r\n* **Alpha Discovery** - finds names nobody else mentioned\r\n* **Autonomous Execution** - converts signals to sized trades\r\n* **CIO** - synthesises all prior layers, weighted by Darwinian agent scores, makes the final call\r\n\r\n---\r\n\r\n## The Autoresearch Loop\r\n\r\nInspired by [Karpathy's autoresearch](https:\u002F\u002Fgithub.com\u002Fkarpathy\u002Fautoresearch). Same pattern, different domain.\r\n\r\n* System identifies worst agent by rolling Sharpe\r\n* Generates one targeted prompt modification\r\n* Runs for 5 trading days\r\n* Checks if agent's Sharpe improved\r\n* Keep (git commit) or revert (git reset)\r\n\r\nThe agent prompts are the weights being optimised. Each trading day is one training iteration. A $20\u002Fmonth VM replaces the H100.\r\n\r\n**Darwinian Weights:** Each agent has a weight between 0.3 (minimum) and 2.5 (maximum). Top quartile agents get weight × 1.05 daily. Bottom quartile get × 0.95. The CIO proportionally weights input by these scores. Good agents get louder. Bad agents get quieter.\r\n\r\n---\r\n\r\n## 18-Month Backtest Results\r\n\r\n**Period:** September 2024 - March 2026 (378 trading days)\r\n\r\n* Prompt modifications attempted: 54\r\n* Survived (kept): 16 (30%)\r\n* Reverted: 37 (70%)\r\n* Deployment phase return: **+22% in 173 days**\r\n* Best individual pick: **AVGO at $152, held for +128%**\r\n\r\nThe system independently discovered its own portfolio manager (CIO) was its weakest component — downweighting it to minimum before we diagnosed the same issue manually.\r\n\r\n### Equity Curve\r\n\r\n![Equity Curve](results\u002Fequity_curve.png)\r\n\r\n---\r\n\r\n## Agent Spawning\r\n\r\nThe system detects recurring knowledge gaps in its own debates. When the same blind spot appears 3+ times in 5 days, it autonomously creates a new specialist agent at neutral weight.\r\n\r\nDuring a 6-month spawning test (Jul-Dec 2024):\r\n- 9 agents spawned autonomously — credit markets, earnings calendar, options flow, liquidity conditions, positioning data, earnings guidance, retail sentiment, technical levels\r\n- 3 went extinct (stuck at minimum weight for 20+ days)\r\n- 6 survived Darwinian selection and reached maximum weight\r\n- Zero human involvement in deciding what to create or when\r\n\r\nThe system grew its own team from 25 to 31 agents based on what it learned it didn't know.\r\n\r\n---\r\n\r\n## All Seasons (PRISM) — Regime-Specific Training\r\n\r\nWe don't train one set of agents and hope they work everywhere. We train separate cohorts on distinct market regimes.\r\n\r\nSame starting agents. Same vanilla prompts. Different evolutionary environments. Completely different survival instincts.\r\n\r\n| Cohort | Period | Return | Modifications Kept | Key Learning |\r\n|--------|--------|--------|--------------------|--------------|\r\n| Bull\u002FLow Vol | 2016-2018 | +7.7% | 180\u002F509 (35%) | Exit vol longs when events resolve peacefully |\r\n| Crisis (COVID) | 2020 Q1-Q2 | -13.1% | 0\u002F3 (0%) | Crashes too fast for autoresearch — agents need to arrive pre-trained |\r\n| Rate Tightening | 2022-2023 | -30.2% | 38\u002F89 (43%) | Don't flip-flop during Fed weeks — 15-day minimum between reversals |\r\n| Recovery | 2020 Q2-Q4 | -29.0% | 0\u002F1 (0%) | Same problem as crisis — too fast for the feedback loop |\r\n| Euphoria | 2021 | +14.3% | 119\u002F243 (49%) | Momentum confirmation before shorts, cap conviction during political crises |\r\n\r\n**Convergent Evolution:** All five cohorts independently discovered the same meta-rules — cap conviction, use VIX as regime filter, enforce hard position limits, never override risk management. Nobody programmed caution. Every cohort learned it from losing money when overconfident.\r\n\r\n**Divergent Evolution:** The same volatility agent started at 844 bytes in every cohort:\r\n- Bull markets: grew to 121,260 bytes (143x). Learned \"exit vol longs immediately when events resolve peacefully.\"\r\n- Rate tightening: grew to 10,354 bytes. Learned \"NEVER buy VXX when VIX is 15-25.\"\r\n- Euphoria: grew to 1,998 bytes. Learned \"require VIX above 30 before going long vol.\"\r\n\r\nSame agent. Same starting prompt. Three completely different survival strategies shaped by three different markets.\r\n\r\n---\r\n\r\n## JANUS Meta-Layer\r\n\r\nMultiple trained cohorts produce different recommendations. JANUS sits above all cohorts and algorithmically weights them by recent accuracy.\r\n\r\nThe weight differential between cohorts is an emergent regime detector:\r\n- When short-window agents outperform → **NOVEL REGIME**\r\n- When long-window agents outperform → **HISTORICAL REGIME**\r\n- When they're roughly equal → **MIXED**\r\n\r\nWe didn't build a regime detector. It emerged from tracking which cohort gets things right.\r\n\r\n---\r\n\r\n## Soros Reflexivity Engine\r\n\r\nMarkets don't just reflect reality — they change it. We built reflexive feedback loops into the simulation framework.\r\n\r\nFive feedback loops modelled:\r\n1. **Price → Fundamentals**: Stock drops >15% trigger credit downgrades, talent flight, capex cuts. Rises >20% trigger cheap capital, talent attraction, customer confidence.\r\n2. **P&L → Behaviour**: Fund drawdown >10% → forced selling cascade. Gains >15% → increased position sizes and concentrated bets.\r\n3. **Narrative → Flows**: 3+ analysts converge on thesis → retail flow follows. Contrarian narratives emerge after extended consensus.\r\n4. **Market → Policy**: Equity drawdown >15% → central bank signals easing. Oil >$130 → strategic reserve releases.\r\n5. **Reflexive Reversal Detection**: Feedback loop running 5+ rounds in one direction → flagged as reflexive extreme. Maximum consensus = maximum fragility.\r\n\r\nFirst detection: Gold bullish consensus appeared in 4 of 5 simulation rounds — flagged as a crowded trade with 32% reversal probability. That's Soros in code.\r\n\r\n---\r\n\r\n## MiroFish Swarm Simulation Integration\r\n\r\nIntegrated with [MiroFish](https:\u002F\u002Fgithub.com\u002F666ghj\u002FMiroFish), a swarm intelligence engine that generates parallel digital worlds populated by thousands of AI agents.\r\n\r\nOur trading agents don't just learn from the past — they train on simulated futures:\r\n\r\n- Overnight, the system generates branching scenarios (geopolitical escalation, Fed policy, earnings shocks, black swans)\r\n- Thousands of simulated agents (fund managers, central bankers, retail traders, corporate executives) interact with reflexive feedback\r\n- ATLAS trading agents are trained inside these simulated futures using the same Darwinian loop\r\n- Agents that navigate simulated futures well get upweighted\r\n- Predictions scored against actual outcomes to improve simulation accuracy\r\n\r\n**First result:** Druckenmiller-style agent scores 1.0 in simulated crashes but 0.22 in melt-ups. Quality compounder agent is the opposite. The system knows which instincts to trust before the regime arrives — not after.\r\n\r\n---\r\n\r\n## Key Insight\r\n\r\nThe orchestration layer matters as much as the intelligence layer.\r\n\r\nIndividual agents improved measurably through autoresearch. But portfolio returns depend on how agent signals are converted to sized positions. The synthesis\u002Fdecision layer is the bottleneck. Improving individual agent intelligence without improving orchestration yields diminishing returns.\r\n\r\n---\r\n\r\n## What's Included\r\n\r\n* Framework architecture and pipeline structure\r\n* Autoresearch loop design\r\n* Backtest results and equity curve\r\n* All Seasons (PRISM) methodology and results\r\n* Agent spawning mechanism\r\n* JANUS meta-layer design\r\n* Soros reflexivity engine\r\n* MiroFish integration bridge\r\n* Example placeholder prompts (generic, not trained)\r\n\r\n## What's NOT Included\r\n\r\n* Trained agent prompts (proprietary — evolutionary products of market feedback)\r\n* PRISM evolved prompts per regime\r\n* CIO active management rules\r\n* Agent scorecard data\r\n* Live portfolio positions\r\n* Darwinian weight values\r\n* MiroFish simulation outputs\r\n\r\nThe trained prompts are the core IP. A competitor starting today is hundreds of iterations behind. That gap widens every day.\r\n\r\n---\r\n\r\n## Tech Stack\r\n\r\n* **Agents:** Claude Sonnet (Anthropic API)\r\n* **Simulation:** MiroFish swarm engine\r\n* **Data:** FMP, Finnhub, Polygon, FRED\r\n* **Infrastructure:** Azure VM ($20\u002Fmonth)\r\n* **Version Control:** Git feature branches for autoresearch tracking\r\n* **Cost:** ~$50-80 for full 18-month backtest, ~$30 for all five PRISM cohorts\r\n\r\n---\r\n\r\n## Contact\r\n\r\n**Chris Worsey** — CEO & Technical Founder, General Intelligence Capital\r\n\r\n[chris@generalintelligencecapital.com](mailto:chris@generalintelligencecapital.com)\r\n\r\n[generalintelligencecapital.com](https:\u002F\u002Fgeneralintelligencecapital.com)\r\n","ATLAS 是一个由 General Intelligence Capital 开发的自改进AI交易代理平台，它结合了 Karpathy 的自动研究方法、索罗斯的反射性理论以及 MiroFish 群体模拟技术应用于金融市场。其核心功能包括通过市场反馈自我优化提示、跨不同市场环境训练、在检测到知识缺口时创建新代理，并模拟未来以应对即将到来的变化。该平台适合希望提高交易策略效率和准确性的投资者使用，支持用户复制现有代理进行实时信号跟踪与自动执行，或根据个人策略构建并测试新的交易模型。此外，用户还可以将自己开发的代理发布到市场中供他人复制使用并获得收益。","2026-06-11 03:49:30","high_star"]