Behind those merges were more than a thousand development commits across dozens of engineering loops.
This cycle wasn't about adding features — it was about building the infrastructure that enables the next generation of Aether products. Major backend architecture improvements, large portions of UX/UI rebuilt, expanded orchestration — all pointed at the same foundation.
01 · Product☁️ Aether Cloud
Aether Cloud received one of its biggest updates to date — 270 commits and 58 merged PRs on main in two weeks.
New models
- Sonnet 5 and Opus 4.8 (1M context) are now live across the platform, with UVT billing repriced to match.
- Neo and Kronus continue receiving major tuning, including new orchestrator skill integrations.
CodePro evolution
The biggest architectural change this cycle was the introduction of Aether System 2. CodePro is evolving from a coding assistant into an engineering orchestration platform: the System 2 compiler lets coordinated agents reason through one shared intermediate representation, enabling long-running engineering sessions instead of isolated prompts. Serenea + Arbiter — our dedicated breaker/builder agent pair — is being introduced for the hardest software engineering tasks, with System 2 risk flags threaded directly into the Arbiter.
- Shared agent IR transcripts
- MPO chaining for longer coherent reasoning
- Unlimited context + local overflow memory
- Quantized TurboVec retrieval
- Atlas truth grounding and hallucination reduction
- Prompt compression through System 2
- Large-session orchestration
The Coder IDE also took a major step toward desktop-class parity: Source Control panel with diffs, real find-in-files, editor settings, a contribution heatmap, and Live Canvas cards rendered directly inside the IDE.
🤖 Aether Agent
Aether Agent shipped a major release — 58 commits on main, dominated by two adversarial security-hardening campaigns and a terminal overhaul.
- Security hardening at depth — 33 verifier-confirmed findings fixed in one campaign alone: path confinement now canonicalizes before comparing, git operations are injection-proof, credential-bearing tokens are checked against safe-URL rules before ever leaving the process, and session logs stopped persisting unredacted task data.
- Terminal experience rebuilt — persisted history, tab completion, a thinking pulse that kills dead air, Ctrl+C that cancels the turn instead of the session, and actionable hints under every failed turn.
- Reliability — stream idle-timeout hardening, honest exit codes (server errors no longer laundered into success), and a first-ever CI pipeline.
- Typed tools, permissions honesty, safer config writes, memory-prune guards — plus a large dead-code removal pass (net −400 lines while adding features).
🌐 Aether Web Platform
Across Aether Cloud and our web applications we focused heavily on polish.
- Unified /platform account hub on the v2 glass design system, with a new agent avatar picker and contribution-graph profile cards.
- Live Canvas gained visible presence, applied-edit flashes, global hotkeys, and plain-language triggers.
- Agent memory UX overhaul — a new agentic-memory write path with user-facing memory preference controls.
- First-run onboarding tutorial — a fires-once guided tour wired through the login funnel.
- Vault sharing — read-only browse, tiered share controls, and a
/sendcommand, all with security scoping verified. - Dozens of UX, accessibility, and reliability improvements across billing, API keys, and device approval.
Agents can now dynamically create new skills during operation. Whether those skills become permanent is determined by our overnight Agent Dreams system — successful behaviors persist while ineffective ones naturally disappear.
🍽 MealPlanner
MealPlanner received focused backend modernization: Google sign-in shipped end-to-end with hardened OAuth callbacks, the CI quality gate was restored as a required check on every PR, and backend security hardening (headers, env validation, host-header guards) is now enforced with matching test coverage. Faster, cleaner, easier to get into — with a larger UX wave staged on feature branches for the next cycle.
02 · ResearchAtlas Preview
This week also marks the first substantial public look at the technology powering the future of Aether. Atlas is not another chatbot. It is the infrastructure we're building to train, coordinate, verify, and continually improve future AI systems — 77 commits and 20 merged PRs this cycle.
OSC — Quantum-Optimized Skill Conditioning
A routing engine that learns which model, skill, prompt, and context produce the best outcome for a task.
- Learns routing decisions instead of stuffing context windows
- Improves across an ever-growing capability atlas
AQRC — Adaptive Quantum Reconstruction Compiler
The dominant research effort of this cycle. Reconstructs computational paths from trained networks and compiles them into increasingly efficient execution geometries.
- Deterministic classical compiler
- Optional quantum optimization — hardware-gated behind a structural-hardness check
- Live Observatory control plane streaming every compilation run
Software first. Quantum only when it proves itself.
Atlas Oracle
Maintains verified truth through provenance, validation, lifecycle tracking, and operator gating. This cycle added a computational drift detector — AQRC helping Atlas catch hallucinations.
Truth flows downward. Reasoning flows upward. Nothing writes truth automatically.
Predator
Predator is becoming our next-generation security research and model-learning pipeline — combining exploit analysis, compiler reasoning, and future MoE routing into one continuously improving system. Upcoming versions integrate GLM 5.2 and DeepSeek V4 Pro fine-tuning.
A note on how we do research. A large share of this cycle's Atlas commits record negative results — failed gates, overturned claims, retracted metrics. Failed experiments remain part of the repository. When a result doesn't replicate on honest re-runs, we overturn it in writing. That discipline is the product.
📈 ATS 2.0 — live and operational
ATS 2.0 is now live and operational: an autonomous AI reasoning freely over live charts, setting its own sleep and wake rhythm, and — in its first live run — learning from the market without taking a single trade. Its doctrine: "No stop, no trade. When uncertain, observe." The research focus is LLM cognition over extended trading windows, powered by our unlimited-context engine. Discipline before capital.
ATS 1.0 is going open source — bring your own Claude, ChatGPT, or Aether API key and run it freely. ATS 2.0 remains exclusive to Aether. Full release post coming separately.
04 · AheadLooking ahead: Aether Q
One of the most exciting projects under development is Aether Q — the first public implementation of AQRC as a usable inference system. Think of it like a memory for computation itself. Instead of solving every request from scratch, Aether Q learns reusable computational paths — the way you stop sounding out a word once you've read it a hundred times. Known problems execute through those optimized paths at a fraction of traditional inference cost; everything else falls back to normal LLM inference. The result is a system that gets faster, cheaper, and more capable the more it's used.
Our goal isn't just to make AI larger. It's to make AI more efficient.
Our direction
- We're investing heavily into Neo.
- We're investing heavily into Kronus.
- We're building Atlas as the foundation that trains future orchestration systems, routing engines, model weights, and agent intelligence.
- We're continuing our quantum research — with a very clear philosophy: software first. Every quantum component must outperform classical baselines before it earns a place in production.
No hype. No magic. Just measurable improvements.
This has been one of the most productive engineering cycles in Aether's history. Much of what shipped this week isn't immediately visible — but it lays the foundation for everything coming next. Atlas, Predator, Aether Q, Neo, Kronus, and the systems behind them are all moving toward the same goal: AI that is more capable, more efficient, and grounded in measurable engineering rather than hype.
We're not just building another model. We're building the systems that help models learn, reason, and improve over time.
We're just getting started.