Proto-AGI harness
Researchers never fully agreed on AGI. They did agree the remaining holes are mostly memory, novel reasoning, and long-horizon agency. Aether wraps a frontier model in a cognitive lattice aimed at those holes — then scores the result against the definitions they actually published.
GPT-5, scored by Hendrycks et al. (2025): 57% of a well-educated adult. Long-term memory storage: 0.
CHC profile
hypothesis2025 · Hendrycks, Song, Szegedy, Bengio, Schmidt, Marcus, Tegmark et al.
A score, not a vibe. GPT-4 27%, GPT-5 57%. The remaining hole is foundational machinery — especially long-term memory storage at 0.
2023–24 · Morris et al., Google DeepMind
Frontier chat models sit at Emerging–Competent AGI: they cover a wide range of tasks at or below a skilled adult, not Expert across the board.
2007 · Shane Legg & Marcus Hutter
Generality is environmental, not benchmark-shaped. A language model in a chat box is one environment.
Where models still fail
| Ability | GPT-4 | GPT-5 | Scaffoldable | The hole |
|---|---|---|---|---|
| General Knowledge | 80 | 90 | medium | Parametric knowledge is wide but stale, uneven, and unowned by a particular life. |
| Reading & Writing | 60 | 100 | low | Frontier models already match or exceed a well-educated adult on most literacy tasks. |
| Mathematical Ability | 40 | 100 | medium | Symbolic slips persist when the model does arithmetic in prose. |
| On-the-Spot Reasoning | 0 | 70 | high | Fluid intelligence on unseen puzzles (ARC-style) and multi-step plans still collapse without search. |
| Working Memory | 20 | 40 | high | Context windows are a poor working memory: they rot, they are not addressable, they are not operated on. |
| Long-Term Memory Storage | 0 | 0 | high | GPT-5 still scores 0. Weights do not update from a conversation. Every session is amnesia. |
| Long-Term Memory Retrieval | 40 | 40 | high | Hallucination is a retrieval-precision failure. Search tools hide it without fixing it. |
| Visual Processing | 0 | 40 | low | Scaffolding in text cannot see. Native multimodal models closed some of this; GUIs and spatial scans remain weak. |
| Auditory Processing | 0 | 60 | low | Speech I/O is now a product feature; rhythmic and musical judgment are not. |
| Speed | 30 | 30 | medium | Token generation is fast at easy tasks and slow at hard ones; it is not human reaction time. |
Hendrycks et al. warn that context windows and RAG are capability contortions: they mask the absence of genuine long-term storage. Aether treats that warning as the experiment, not as a reason to stop. The battery tells you whether the prosthesis held.
The lattice
By 2026 the field had quietly converged: not a swarm of specialist agents, but a thin ReAct loop wrapped in memory, tools, planning, and skills. Claude Code, Deep Agents, OpenHands, Letta — different products, the same skeleton. Aether is that skeleton pointed at AGI definitions instead of at git.
Episodic memory
Letta, Mem0, Hermes
Write what happened. The prosthetic for the CHC zero — long-term storage.
Semantic memory
Mem0, Graphiti/Zep
Facts about you, the world, and this project, stored as durable claims.
Working memory
SWE-agent ACI, Deep Agents virtual FS
A scratchpad the model can edit instead of stuffing the context window.
Hierarchical planner
Deep Agents, OpenHands, Tree of Thoughts
Sketch a plan before acting. System-2 as a first class step.
Reflexion critic
Shinn et al. Reflexion
A second pass that hunts confabulation and plan holes.
Tool use
ReAct, MCP, Claude Code
Calculator, clock, and memory tools. Exactness the weights will not guarantee.
World model
Goertzel architectures, belief state
A living note of what is believed to be true right now.
Metacognition
Burnell et al. coverage gap
Tag confidence. Say 'I don't know' when the store is empty.
Skill library
Voyager, Anthropic Skills
Save a procedure once, reuse it. Tiny skill-acquisition loop.
Internal debate
CrewAI, AutoGen, multi-agent debate
Two perspectives, one decision. Off by default — it costs a call.
Goal stack
AutoGPT, Legg-Hutter agency
Track standing goals so turns are not amnesiac tasks.
Constitution
Constitutional AI, CLAUDE.md
Standing instructions. Identity as a file, not a mood.
Harness atlas
Anthropic, 2025
The result that flipped the field: a thin ReAct loop with files, bash, grep, CLAUDE.md memory, and compaction beat elaborate multi-agent graphs on real software work.
LangChain, 2025–26
The batteries-included Claude Code analog on LangGraph: planning tool, virtual filesystem, sub-agents, summarization, skills.
LangChain, v1 2025
Production default for long-horizon work: explicit state, checkpoints, replay, human-in-the-loop. Ranked first for durable tasks in 2026 surveys.
All Hands AI (ex-OpenDevin)
Self-hostable autonomous engineer. Plans a dependency graph, then fans work out to sub-agents in a sandbox.
Princeton + Stanford, NeurIPS 2024
Defined the Agent-Computer Interface. mini-swe-agent later showed ~100 lines can still clear 74%+ on SWE-bench Verified.
Berkeley / Letta
Treats context as RAM and a vector/archival store as disk. The agent pages memories in and out.
That is the entire MS experiment, the one GPT-5 still scores at zero. If the lattice can hold a nonce across a turn, scaffolding is doing real work. If it cannot, we learned something cheaper than a manifesto.
Create a chamber