True autonomous agents are 'not anywhere close' — AGI and reliable long-horizon agents will require a full decade (2034 or beyond) to develop the holistic contextual reasoning and robust world models needed for unconstrained physical and digital enviro...

True autonomous agents are 'not anywhere close' — AGI and reliable long-horizon agents will require a full decade (2034 or beyond) to develop the holistic contextual reasoning and robust world models needed for unconstrained physical and digital environments. | Long-horizon agent benchmark breakthrough

Key catalyst: Long-horizon agent benchmark breakthrough

Watch events: Long-horizon agent benchmarks; world-model saturation

Karpathy "Software 3.0" framing (INF_026) validated; specific decade-out AGI timeline provides bearish anchor against Altman 2030 / Aschenbrenner 2027.

Adverse (4)

{
  "nia": false,
  "mode": "FORECAST",
  "role": "Cited-Other",
  "context": "Karpathy's structural-caution anchor provides the bearish extreme in the AGI-timeline debate. Couples with Hassabis jagged-intelligence (AI_005) and CMQ_047 (agents close-the-loop without human).",
  "to_year": 2034,
  "conv_cues": "decade horizon; ex-senior-researcher framing",
  "direction": "HAPPEN",
  "from_year": 2034,
  "timeframe": "2034+",
  "conv_level": "HIGH",
  "milestones": [
    {
      "kind": "llm_pre_event",
      "label": "Karpathy publishes update on his decade-away framing or releases follow-up agentic research (Eureka Labs, nanochat successor)",
      "source": "https://fortune.com/2026/03/17/andrej-karpathy-loop-autonomous-ai-agents-future/",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -10,
      "source_id": null,
      "confidence": 0.85,
      "expected_date": "2027-05-16",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2028-06-30",
        "from": "2026-04-01"
      },
      "measurement_criterion": "Karpathy public talk, blog, or repo release where he addresses his decade timeline, allowing direct comparison to Oct 2025 Dwarkesh interview"
    },
    {
      "kind": "llm_pre_event",
      "label": "Continual-learning architecture demonstrated in production agent (memory persists across sessions, no model retraining)",
      "source": "https://www.predictiveanalyticsworld.com/machinelearningtimes/agi-is-still-a-decade-away-todays-ai-agents-are-slop-openai-cofounder-andrej-karpathy/13949/",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -9,
      "source_id": null,
      "confidence": 0.5,
      "expected_date": "2027-12-15",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2029-06-30",
        "from": "2026-06-01"
      },
      "measurement_criterion": "Frontier lab paper or product demo shows agent that genuinely accumulates and uses session-spanning learning without full retraining cycle"
    },
    {
      "kind": "llm_pre_event",
      "label": "Frontier agentic benchmark (OSWorld / GAIA / AgentBench long-horizon) crosses 70% pass-rate",
      "source": "https://www.remio.ai/post/why-andrej-karpathy-says-ai-agents-are-a-decade-from-reality",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -8,
      "source_id": null,
      "confidence": 0.6,
      "expected_date": "2028-03-16",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2029-12-31",
        "from": "2026-06-01"
      },
      "measurement_criterion": "Public leaderboard for recognized long-horizon agentic benchmark shows >=70% (vs 2025 SOTA typically 30-50% on hardest variants)"
    },
    {
      "kind": "llm_pre_event",
      "label": "Frontier lab (OpenAI, Anthropic, DeepMind) ships agent product with persistent multi-day autonomous task completion at >85% reliability",
      "source": "https://www.landera.ai/guide/karpathy-paradox",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -7,
      "source_id": null,
      "confidence": 0.55,
      "expected_date": "2029-07-01",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2031-12-31",
        "from": "2027-01-01"
      },
      "measurement_criterion": "Lab product release with marketing claim of multi-day autonomous task completion plus published evaluation showing >85% success on long-horizon real-world task suite"
    },
    {
      "kind": "llm_pre_event",
      "label": "Multi-modal robotic agent demonstrates >24-hour unscripted task continuity in real-world environment",
      "source": "https://www.flowhunt.io/blog/the-decade-of-ai-agents-andrej-karpathy-agi-timeline/",
      "status": "pending",
      "weight": 0.4,
      "ordinal": -6,
      "source_id": null,
      "confidence": 0.5,
      "expected_date": "2030-07-02",
      "research_origin": "training",
      "expected_date_range": {
        "
... (truncated)