Aether Systems · Engineering Journal

Inside AetherATS 2.0: Teaching an AI When Not to Act

July 15, 2026 — Notes from the first live operational windows of our long-horizon cognition experiment. The hardest trade is the one you never take.

The first thing ATS 2.0 learned wasn't how to trade. It learned when not to.

AetherATS v2 operator console: account balance, intervention controls, live TradingView chart, and autonomous run controls — ATS 2.0 reasoning over live market data while remaining fully operator-observable.

During one of its first live sessions, ATS continuously monitored multiple markets. It analyzed. It waited. It refused to invent certainty.

It executed zero trades. That wasn't failure. That was exactly what we designed it to do.

Why that matters

Most AI trading systems optimize for activity. ATS optimizes for confidence. If conviction isn't there, it waits — and it can tell you exactly why.

ATS live reasoning log showing DATA NOT READY entries and sleeping/auto-wake behavior, next to a 5/5 system health check panel — The reasoning log during degraded data conditions, beside a live 5/5 system healthcheck.

One of my favorite log entries from these first sessions reads:

[DATA NOT READY] NQ=F: missing vision (have: chart, options). NOT reasoning a trade — that would mean deciding on data I don't have. Sleeping ~120s; auto-wakes when the feeds come alive.

Most AI demos only show successes. This shows restraint. The system noticed its own data feeds were degraded, declined to reason about a trade it couldn't ground, put itself to sleep, and scheduled its own wake-up. No human told it to do any of that.

That tells you more than a performance chart ever could.

The cognition loop

ATS runs continuously through a six-stage cycle:

Observe↓ Recall↓ Reason↓ Ground↓ Propose↓ Record

Every cycle updates persistent memory. Nothing is forgotten between restarts.

Persistent cognition

ATS keeps a 5 GB persistent reasoning space that survives restarts — built on the same unlimited-context engine behind CodePro. Not 5 GB of prompts. Persistent cognition. It remembers:

why it entered positions
why it avoided them
what it learned from each window
which of its own assumptions failed

Understanding compounds instead of resetting.

The operator console

The interface isn't a dashboard — it's an operator console. Four concepts, all visible in the screenshots above:

Autonomous executionGo Live / Pause / Stop — the agent runs on its own schedule.

Human interventionBuy / Flatten / Size — an operator can step in at any moment.

Reasoning transparencyLive logs stream every decision, including refusals to act.

System healthMemory, feeds, tools, and context verified continuously.

Every decision the agent makes is observable. Operators can inspect its reasoning loop, monitor system health, intervene immediately, or let the system continue autonomously. Building trustworthy autonomous systems means making them inspectable, not opaque.

The philosophy

"No stop, no trade.
When uncertain, observe."

ATS values correctly avoided mistakes as much as successful trades. A caught trap counts as a win. Its retune notes — records of its own uncertainty and rejected reasoning — are treated as its most valuable output.

Why we're building this

ATS isn't simply a trading system. It's one of the environments where we're studying long-duration AI cognition: what does an AI learn after watching a live environment for days instead of minutes?

Markets provide immediate feedback. Every decision is measurable. Every mistake costs something. That makes them an excellent research environment — and it ties directly into the rest of the stack:

AtlasATS feeds observations into Atlas; Atlas grounds ATS's reasoning against verified truth. AQRCwill eventually compile ATS's learned reasoning paths into faster execution geometries. OSCwill eventually route ATS's tasks to the best model, skill, and context for each situation.

The trading system is the testbed. The cognition stack is the product.

ATS 1.0 goes open source

The original Aether Trading System will be open source. Bring your own Claude. Bring your own GPT. Bring your own Aether.

ATS 2.0 remains internal while we continue researching long-horizon reasoning.

The bigger vision

Markets happen to be one of the most demanding real-time environments an AI can experience. Every decision has consequences. Every mistake has a measurable cost. Every good decision can be evaluated against reality.

That's why ATS exists. It's not just about building a better trading system — it's about understanding how an AI learns over time.