A wide, opinionated sweep of the lattices people actually ship: coding agents, memory OS-es, runtimes, protocols, and the three prompt patterns everything else is made of. Each entry is tagged with the AGI gap it is quietly trying to close.
Claude Code / Claude Agent SDK
Anthropic, 2025
productThe 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.
- Mechanism
- Single-threaded tool loop, permissioned diffs, context compaction, project memory file.
- AGI gap
- Working memory (compaction), retrieval (CLAUDE.md), reasoning (the loop itself).
- Lesson for Aether
- Harness quality over graph topology. Aether's chamber is this shape: one loop, many layers as tools.
Deep Agents
LangChain, 2025–26
runtimeThe batteries-included Claude Code analog on LangGraph: planning tool, virtual filesystem, sub-agents, summarization, skills.
- Mechanism
- Middleware on an explicit graph state. Harness engineering as the 2026 lever.
- AGI gap
- Planning, WM (virtual FS), procedural skills.
- Lesson for Aether
- Aether's skill library and planner are the portable version of this idea.
LangGraph
LangChain, v1 2025
runtimeProduction default for long-horizon work: explicit state, checkpoints, replay, human-in-the-loop. Ranked first for durable tasks in 2026 surveys.
- Mechanism
- Graph + reducer + checkpointer. Recovery is a first-class primitive.
- AGI gap
- Horizon / autonomy (METR), not a CHC faculty.
- Lesson for Aether
- Aether persists traces and memories so a later turn can resume. Full graph recovery is the next rung.
OpenHands
All Hands AI (ex-OpenDevin)
productSelf-hostable autonomous engineer. Plans a dependency graph, then fans work out to sub-agents in a sandbox.
- Mechanism
- Planner → DAG of tasks → sandboxed execution.
- AGI gap
- Planning, tools, speed of software labor.
- Lesson for Aether
- Hierarchy matters once a task has real dependencies. Aether's planner is the single-agent form.
SWE-agent / mini-swe-agent
Princeton + Stanford, NeurIPS 2024
evalDefined the Agent-Computer Interface. mini-swe-agent later showed ~100 lines can still clear 74%+ on SWE-bench Verified.
- Mechanism
- Constrained edit/command interface; issue in, patch out.
- AGI gap
- Reasoning on software environments; WM via editor state.
- Lesson for Aether
- The interface is the intelligence. Aether's tools are a small ACI: memory, calc, world, skills.
Letta (MemGPT)
Berkeley / Letta
memoryTreats context as RAM and a vector/archival store as disk. The agent pages memories in and out.
- Mechanism
- Tiered memory with explicit page-in/page-out tools.
- AGI gap
- MS and MR — the CHC zeros.
- Lesson for Aether
- This is the load-bearing prosthetic. Aether's episodic/semantic stores are a Letta-shaped design.
Drop-in memory layer extracting, consolidating, and retrieving user facts across sessions.
- Mechanism
- Extraction + dedup + retrieval API.
- AGI gap
- MS / MR for personalization.
- Lesson for Aether
- Memory must be extracted, not merely logged. Aether writes memories as a first-class tool, not a transcript dump.
Hermes Agent
Nous Research
productPersonal agent that 'grows with you' — always-on, self-improving, local-leaning.
- Mechanism
- Persistent identity + skill accumulation over calendar time.
- AGI gap
- MS as a life, not a session. Goals.
- Lesson for Aether
- Proto-AGI has to be the same someone tomorrow. Aether binds memory to an account.
Role-playing multi-agent crews. Fastest path to a researcher/writer/critic split.
- Mechanism
- Role, goal, backstory, sequential or hierarchical process.
- AGI gap
- Reasoning via specialized perspectives (debate layer).
- Lesson for Aether
- Aether's debate layer is a two-voice crew without the orchestration tax.
AutoGen → Microsoft Agent Framework
Microsoft
runtimeConversational multi-agent, then folded into MAF 1.0 with Semantic Kernel's enterprise runtime. AutoGen itself is in maintenance.
- Mechanism
- Agent chat, code sandbox, handoffs.
- AGI gap
- Reasoning, tools, recovery.
- Lesson for Aether
- The industry moved from conversation graphs to runtimes. Aether is a runtime with one occupant.
OpenAI Agents SDK / Codex
OpenAI
productManaged agents plus a coding product collapsing into ChatGPT. Computer-use as embodiment-lite.
- Mechanism
- Hosted tools, computer use, handoff patterns.
- AGI gap
- Vision (screens), tools, economic work.
- Lesson for Aether
- Embodiment in this era is a pointer and a shell. Aether stays text-native and admits the vision gap.
Agent Development Kit: orchestrated agents with Gemini, first-class tools and evaluation.
- Mechanism
- Declarative agents + tool surface + eval hooks.
- AGI gap
- Evaluation-driven generality.
- Lesson for Aether
- A harness without a battery is a chatbot. Aether ships the battery next to the loop.
OpenCode / Cline / Roo / Aider / Goose / Crush
2024–26 terminal agents
productThe terminal-coding cluster. Plan-then-act (Cline), repo maps (Aider), local-first (Goose/Block), CLI-native (OpenCode, Crush, Gemini CLI).
- Mechanism
- Repo as environment. Diffs as the write primitive. User approval as a value.
- AGI gap
- Tools, WM, horizon on a codebase.
- Lesson for Aether
- Most 'agent' products are ACIs over git. Aether is an ACI over a mind: memory, goals, skills, world.
Voyager
MineDojo / NVIDIA, 2023
patternOpen-ended Minecraft agent that writes, stores, and reuses skills. Curriculum from automatic novelty.
- Mechanism
- Skill library + iterative prompting + embodied environment.
- AGI gap
- Procedural memory, adaptation, Chollet-style skill acquisition in one world.
- Lesson for Aether
- Aether's skill store is Voyager without the game. Skills must be reused, not just saved.
ReAct / Reflexion / Tree of Thoughts
Yao, Shinn, Princeton / DeepMind, 2022–23
patternThe three primitive lattices: interleave reason+act; verbal self-critique into the next try; search a tree of plans.
- Mechanism
- Prompt-level control flow. No new weights.
- AGI gap
- On-the-spot reasoning.
- Lesson for Aether
- Aether implements all three as toggleable layers. That is the experiment: how much of R is search.
MCP + Anthropic Skills
Anthropic, 2024–26
protocolModel Context Protocol standardizes tools. Skills package procedures the agent can load. Superpowers-style methodologies spread the same idea.
- Mechanism
- Typed tools + on-demand playbooks.
- AGI gap
- Tools and procedural knowledge.
- Lesson for Aether
- Aether's tools and skill library are a private MCP. The protocol is how lattices compose.
The first public 'loop until the goal is done' agents. Recursion without recovery. Historically important, operationally fragile.
- Mechanism
- Goal → tasks → execute → new tasks.
- AGI gap
- Agency / goals. Demonstrated the failure mode of unconstrained loops.
- Lesson for Aether
- Aether keeps a goal stack and a budget (one user turn, bounded tool rounds). Horizon without a halt is a bill.
Compile prompts and tool use against a metric. Treat the lattice as a program to optimize.
- Mechanism
- Signatures, modules, teleprompters.
- AGI gap
- Adaptation — learning the scaffold itself.
- Lesson for Aether
- Aether's battery is the metric. We do not yet compile the layers; that is the obvious next experiment.
OpenClaw / DeerFlow / OpenHuman
2025–26 personal SuperAgents
productAlways-on personal agents (OpenClaw), long-horizon research SuperAgents (ByteDance DeerFlow on LangGraph), local-first memory companions (OpenHuman).
- Mechanism
- Presence + memory + research tools, with different discipline/safety postures.
- AGI gap
- MS over calendar time; METR horizon on research.
- Lesson for Aether
- Proto-AGI is lived-with, not demoed. Aether's login exists so the lattice can be a companion, not a pastebin.