1. 🧩 Three Posts, One Synthesis
This site has now made the AGI argument three times, each time adding a layer. The AGI Threshold argued AGI isn't a single switch — it's a 9-layer capability stack filled in unevenly. Astra, Claude & Gemini put real, named models and sourced benchmark claims onto that argument, and surfaced the ARC-AGI-3 harness controversy as live proof that benchmark numbers alone can't settle the question. The Six Levers asked a different question entirely — not "how capable," but "what's actually driving the next advance" — and found pre-training scale running low on runway while agentic scaffolding and test-time compute do most of the recent work.
This post fuses all three: which lever moves which stack layer, which of those layer-lever pairs are close to solved versus genuinely open, and what that means concretely for a researcher, a builder, and an investor reading this in September 2026.
9 LayersThe capability stack — unchanged from the original framework, now cross-referenced against six levers and three research-readiness tiers
3 of 9Stack layers we classify as near-term/engineering-solvable within ~12 months using current techniques
3 of 9Layers we classify as open research — memory, metacognition, safe autonomy — with no agreed solution path yet
49 GWProjected 2028 US power shortfall — the constraint that applies no matter which layer you're trying to close
2. 🏗️ The Capability Stack, in Full
Same nine layers as the original framework — reasoning at the base, safe autonomy at the top — now annotated with the lever (from our six-lever framework) most responsible for advancing each one, and the research-readiness tier (Section 5) it currently sits in.
Lever 1: Pre-training/Data
Lever 3: Test-Time Compute
Lever 4: Post-Training
Lever 5: Agentic Scaffolding
Lever 6: Memory/Continual Learning
🛡️
Safe Autonomy
Knowing when to stop, escalate, or ask rather than act.
🪞
Metacognition
Calibrated self-awareness — knowing what it doesn't know.
🔁
Reliability & Self-Correction
Staying dependable across long chains of steps.
📅
Long-Horizon Planning
Pursuing a goal across hours or days without losing coherence.
🖱️
Tool & Computer Use
Turning intelligence into action via code, APIs, interfaces.
🧠
Memory & Continual Learning
Learning from experience without retraining from scratch.
🌐
Multimodality → World Coherence
Text, image, audio, video as one coherent world model.
🔀
Generalization
Transferring reasoning into genuinely unfamiliar situations.
💡
Reasoning
Reliably solving novel, complex problems.
Read this diagram top to bottom and the pattern from all three prior posts recurs one more time: the layers with an assigned lever that's cheap and fast (test-time compute, post-training, scaffolding) are near-term; the layers where the only plausible lever is memory/continual-learning research are open. The stack hasn't changed — what's new here is that every layer now has an accountable "who's supposed to fix this" answer, and for three of nine layers, the honest answer is "nobody has a working recipe yet."
3. 📍 Placing the Frontier Models on the Stack
Using the sourced claims from our Astra/Claude/Gemini post, here's where today's named frontier models actually sit — not a leaderboard, a placement.
| Stack Layer | GPT-6 Astra | Claude Opus 5 | Claude Fable 5.1 | Gemini 3 Deep Think | Open-Weight Tier |
|---|
| Reasoning | 62.7%* | 30.2% | n/a public | disputed 45–85% | competitive, cheaper |
| Tool & Computer Use | OSWorld 72.6% | permission-first | long-run agent focus | Antigravity platform | improving fast |
| Long-Horizon Autonomy | Portal in 24h | 14h30m record (4.6) | recommended for multi-day | not independently shown | lagging |
| Transparency/Metacognition | internal concern flagged | no reported regression | no reported regression | not publicly assessed | largely unmeasured |
| Safe Autonomy Architecture | Daybreak gating only | real-time classifiers | real-time classifiers | not publicly detailed | minimal disclosed |
| Sovereign/Cost Efficiency | premium tier | tiered pricing | premium tier | Flash variants cheap | DeepSeek/Qwen strongest here |
*Astra's ARC-AGI-3 standard-harness score; provider-adapter harness reported 99.9%+ — see the Astra/Claude/Gemini post for the full harness discrepancy. 🟢 Strong 🟡 Good-partial 🟠 Partial/disputed 🔴 Weak/undisclosed
No single model leads every row, and the row every model is weakest on is the same one across the board: transparency and safe-autonomy disclosure. That's not a coincidence — it's the direct, visual confirmation of Section 2's stack diagram. The layer with no cheap lever is the layer no model looks strong on.
4. 🔧 Layer by Layer: Which Lever Actually Moves It
💡 Reasoning ← Test-Time Compute
Extended "thinking" before answering is the direct, proven lever here — RLVR-trained reasoning models and inference-time scaling are the mechanism behind essentially every recent reasoning benchmark jump.
🔀 Generalization ← Post-Training
RLVR and on-policy distillation extract more consistent transfer from an already-trained base model — this is squeezing existing capability, not adding new knowledge, which is exactly why it's cheap and fast.
🌐 Multimodality ← Pre-Training/Data
Genuine world coherence across modalities still depends on what's in the training mix — video pretraining (natural + synthetic) is the active frontier here, making this the one near-the-top layer still tied to Lever 1's shrinking runway.
🧠 Memory ← Continual Learning Research
No shipped scaffolding trick substitutes for actually learning from experience — this is Sutskever's explicit bet and MemOS-style memory-architecture research, both pre-product as of this writing.
🖱️ Tool & Computer Use ← Agentic Scaffolding
The cheapest, fastest-moving lever of the six — a harness swap alone produced a 34-point score change on an identical model in Cursor's own research, with zero change to weights.
📅 Long-Horizon Planning ← Scaffolding + Test-Time Compute
A blend: orchestration/retry logic extends coherent operation, while extended reasoning within each step reduces the compounding-error problem — neither lever alone is sufficient.
🔁 Reliability ← Test-Time Compute (Partial)
Self-verification during extended reasoning helps, but doesn't fully solve compounding error over very long step counts — this is why we classify it as a hard problem, not near-term, despite having an active lever.
🪞🛡️ Metacognition & Safe Autonomy ← Memory/Continual Learning Research
Both require the kind of calibrated, experience-updated self-model that no current lever cleanly produces — the honest answer, per Section 6, is that these need new research, not more scaffolding or compute.
5. ⏱️ Classifying the Gaps: Near-Term, Hard Problem, Open Research
This three-tier classification is our editorial synthesis — not a single source's claim — built from where the six-lever framework and the per-model evidence converge.
🟢 Near-Term (~12 months)
Current techniques, known recipe, mostly engineering execution
Reasoning — test-time compute is proven; the remaining work is efficiency (uncertainty-aware allocation at <0.6x naive cost)
Generalization — RLVR + distillation pipelines are mature, widely adopted paradigms already
Tool & Computer Use — scaffolding gains of 30+ points are demonstrated and repeatable; this is the "just build it better" layer
🟡 Hard Problem (Multi-Year)
Known direction, no solved recipe, real engineering + research effort needed
Multimodal World Coherence — depends on video/synthetic data still being built out, and on the pretraining data wall not closing that door first
Long-Horizon Planning — orchestration is improving fast but coherent multi-day autonomy across domains remains unproven at scale
Reliability at Scale — compounding error over long step-chains has no full fix yet, only partial mitigations
🔴 Open Research (Uncertain Horizon)
No agreed solution path; may require new architectures or paradigms entirely
Memory / Continual Learning — Sutskever's explicit "age of research" bet; catastrophic forgetting remains unsolved
Metacognition — calibrated self-awareness has no current training signal that reliably produces it
Safe Autonomy — the OpenAI internal CoT-monitorability dissent (see Astra post) shows even frontier labs are unsure how to instrument this, let alone solve it
6. 🔬 The Path Forward for Researchers
🧩
Continual Learning Without Forgetting
The core unsolved problem separating frozen-between-training-runs models from anything that learns "on the job" — progress here looks like architectures that update durably from experience without erasing prior knowledge.
🔍
Interpretability for Monitorability
The Astra CoT-monitorability dissent is a live research problem right now — finding substitutes for "read the reasoning trace" as reasoning becomes less transparent is urgent, not theoretical.
✅
Verifiable Rewards for Non-Verifiable Domains
RLVR works because math and code have checkable answers. Extending anything like it to judgment-heavy, ambiguous domains (the layers that matter most for safe autonomy) is an open design problem, not an engineering one.
🌍
Grounded World Models
2026 is being called a breakthrough year for early world-model and continual-learning prototypes — but "prototype" is the operative word. Progress here looks like reliability at deployment scale, not just research demos.
7. 🛠️ The Path Forward for Builders
The good news for engineers and startups: you don't need to wait on Section 6's open research to ship real improvements. The near-term tier (Section 5) is a build roadmap available today.
🔌
Invest in the Harness, Not Just the Model
Lever 5
A 34-point benchmark swing from harness redesign alone, with zero model change, is the single highest-leverage finding for any team building on top of frontier models rather than training them.
🧵
Design Around Compounding Error
Layer: Reliability
Since reliability at long horizons is a hard-problem tier, build checkpointing, verification steps, and human-in-the-loop gates into agentic workflows now rather than assuming the model will get there first.
💾
Engineer Memory Workarounds
Layer: Memory (open research)
Since true continual learning is open research, retrieval-augmented and structured-memory workarounds (MemOS-style patterns) are the practical substitute available today — not a permanent fix, but a real one.
📏
Build Evals Before You Build Features
Layer: All
Given how much benchmark numbers vary by harness alone (Section 3's footnote), teams that skip building their own task-specific evals are optimizing against numbers that may not transfer to their actual use case.
8. 💰 The Commercial Angle
💵 Monetizable Today
Coding agents and tool-use products — the layer with the clearest lever (scaffolding), cheapest to iterate on, and already generating real revenue across the industry
Reasoning-as-a-service (API tiers priced by reasoning depth) — Claude Opus 5's near-Fable performance at roughly half the price is a direct example of post-training economics passed to customers
Open-weight/cost-efficient deployment (DeepSeek V4, Qwen3.8-Max, Gemini Flash variants) — the layer where price competition is most active right now
🔬 Pre-Revenue Research Bets
Continual learning and persistent memory architectures — SSI's entire thesis is a research bet with no shipped product yet
Safe, fully autonomous long-horizon agents — the layer every model in Section 3 scores weakest and discloses least about; genuinely autonomous deployment at scale isn't underwritten by insurance or enterprise trust yet
Grid-scale power and compute infrastructure — capital-intensive, multi-year build cycles (Meta's Hyperion/Prometheus campuses) that underwrite every other bet on this list, but generate no direct model-layer revenue themselves
The economic pattern maps cleanly onto the tier classification: near-term layers are cheap to improve and already monetized; open-research layers are exactly where capital is being deployed on long, uncertain bets rather than product cycles. That's a rational allocation, not a coincidence — money follows tractability except where a lab makes a deliberate long-horizon research bet (Sutskever's SSI being the clearest example).
9. 🚧 The Bottleneck That Isn't About Models At All
Every layer, every lever, and every tier above assumes compute is available to spend. It increasingly isn't, cheaply. As covered in The Six Levers, the binding 2026 constraint has shifted from chip availability to grid-scale power — Morgan Stanley projects a roughly 49GW US shortfall by 2028, and high-voltage transformer lead times now stretch to five years. This applies uniformly across the stack: a research breakthrough in continual learning still needs power to train and serve at scale, exactly like a scaffolding improvement needs power to run more agent instances in parallel. No tier in Section 5 is exempt from this constraint — it just hits capital-intensive tiers (pre-training scale, frontier-scale research bets) harder and faster than cheap, inference-light layers like scaffolding improvements.
10. 🗺️ Putting It Together: A Forward-Looking Roadmap
Next 12 Months
Near-Term Layers Consolidate
Reasoning, generalization, and tool/computer use keep improving via cheaper test-time compute and better scaffolding — expect incremental, frequent releases rather than generational leaps, and expect harness engineering to matter as much as model choice.
1–3 Years
Hard Problems Get Partial Fixes
Long-horizon planning and reliability improve meaningfully but not completely — expect longer autonomous task durations (building on the 14h30m → 24h trend) with continued reliability caveats, and multimodal world coherence advancing as video-pretraining techniques mature, racing against the 2028 data-wall projection.
3+ Years / Uncertain
Open Research Either Breaks Through or Doesn't
Memory, metacognition, and safe autonomy depend on genuinely new paradigms (Sutskever's bet, interpretability research, verifiable-reward extensions) succeeding — this tier has no guaranteed timeline, and Dario Amodei's own 1–3 year "powerful AI" prediction window sits inside this uncertainty, not outside it.
The stack doesn't close all at once. It closes lever by lever, tier by tier — and the honest roadmap has three layers still marked "open" with no committed date.
11. 🧭 Verdict
🎯 The Bottom Line
Put the models, the levers, and the timeline together and one picture emerges clearly: the AGI capability stack is not closing uniformly, and it isn't going to. The bottom three layers — reasoning, generalization, tool use — are near-term engineering problems with proven levers, already being monetized, and likely to keep improving on a roughly annual cadence. The middle layers — planning, reliability, multimodal coherence — are hard problems with a known direction but no finished recipe, likely to show real but partial progress over the next few years. The top three — memory, metacognition, safe autonomy — are genuinely open research questions that even frontier labs' own safety teams are visibly unsure how to solve, sitting behind a research bet (continual learning) rather than a scaling recipe. For researchers, that's the actual open frontier. For builders, the near-term tier is a roadmap available right now. For capital, the split between "monetizable this year" and "multi-year research bet" is already visible in where the money is actually going. And underneath all of it, gigawatts — not benchmarks — set the real pace.