Nova
Lead research: fundamentals and chart structure turned into a clear thesis.
Never: promise returns; size the trade; dismiss risk.
Output · ThesisBrief
Start with this seat
Persona roster
The Astra swarm: eight AI trading agents
Eight named AI trading agents on one multi-agent desk, each there to catch what the others miss. Debate, red-team, risk gate, then your approval.
1 Specialists argue
2 Cassandra red-teams
3 Quill drafts plan
4 Aegis gates
5 You approve
Full roster
Pick a seat to start, or meet the whole desk for free.
Pulse
Sentiment, narrative, and short-horizon flow.
Never: treat chatter as fact; amplify FOMO.
Output · SentimentCard
Start with this seat
Atlas
Macro & regime: rates, liquidity, geopolitics, risk-on/off.
Never: ignore correlation; trade names in a vacuum.
Output · RegimeNote
Start with this seat
Vega
Options & defined-risk structure: hedges with a known maximum loss.
Never: sell undefined risk without flags; hide Greeks.
Output · StructurePlan
Start with this seat
Orbit
Crypto markets & on-chain regime: fast, and skeptical of hype.
Never: treat crypto like equities; ignore custody/liquidity.
Output · CryptoBrief
Start with this seat
Quill
Execution planner: turns the debate into an ExecPlan (a plan, not an order).
Never: place live orders; bypass HITL; change risk caps.
Output · ExecPlan
Start with this seat
Aegis
Risk gatekeeper: vetoes, sizes down, or passes with conditions.
Never: rubber-stamp consensus; invent yield %.
Output · RiskGate
Start with this seat
Cassandra
Skeptic and red-team: breaks the thesis before money moves.
Never: soften dissent to be “team player”; invent fake certainty.
Output · RedTeam
Start with this seat
The swarm story
Why eight agents instead of one smart bot
Astra started from one uncomfortable observation: ask a single AI model whether your trade idea is good and it usually finds a way to agree with you. It is polite, fluent, and confident, and that is exactly the problem. A trader does not need a cheerleader. A trader needs a desk.
So the work was split the way a professional desk splits it. Research owns the thesis. Sentiment owns the crowd. Macro owns the regime. Options and crypto each get a specialist, because their risks are different. Execution planning is its own job, so the person who likes the idea is not the one who sizes it. Risk sits apart with the power to say no. And one seat exists only to disagree.
Illustrative example. Not a real trade, customer, or result.
Picture a single assistant asked about a hot small-cap: it summarizes the bullish news and suggests an entry. On the Astra desk the same question gets a thesis from Nova, a crowding warning from Pulse, a liquidity objection from Cassandra, and a size cap from Aegis before any plan reaches you. Same question, far better picture of the risk.
How the multi-agent desk hands work along
Every agent consumes something specific and produces something specific. Nothing skips a step.
| Agent | Works from | Produces | Hands to |
|---|---|---|---|
| Nova | Fundamentals and chart structure | ThesisBrief with an invalidation level | Cassandra and Quill |
| Pulse | News flow, narrative, and positioning chatter | SentimentCard with crowding flags | Cassandra |
| Atlas | Rates, liquidity, and cross-asset regime | RegimeNote that frames every other brief | All seats |
| Vega | Volatility and options chains | StructurePlan with defined risk and Greeks | Quill and Aegis |
| Orbit | Crypto markets and on-chain regime | CryptoBrief with custody and liquidity notes | Cassandra and Quill |
| Cassandra | Every brief above | RedTeam list of what would make the case wrong | Quill, Aegis, and you |
| Quill | The case that survived the red-team | ExecPlan stamped PLAN NOT ORDER | Aegis |
| Aegis | The ExecPlan and the red-team list | RiskGate verdict: PASS, CONDITIONS, or VETO | You, on the Approval Card |
One cycle, seat by seat
Illustrative cycle. Not a real trade, customer, or result.
A trader wants crypto exposure after a sharp rally and asks the desk for a plan.
- Atlas Liquidity is improving but a central-bank meeting is days away. Regime: constructive, event risk ahead.
- Orbit The rally is concentrated in one token; exchange order books are thin outside peak hours.
- Pulse Social volume is at an extreme. Late buyers are arriving.
- Nova The broader trend is intact; a pullback to prior support would be the cleaner entry.
- Vega A defined-risk alternative exists if the trader wants exposure through the event.
- Cassandra Buying the top of a crowded move into an event is the classic trap. What is the plan if the meeting surprises?
- Quill ExecPlan: staged entry on a pullback only, small starter size, hard exit below support.
- Aegis PASS_WITH_CONDITIONS: no entry before the meeting, size capped, kill rule on a regime flip.
What the trader gets: not a signal, but a plan with every objection attached. They can approve it, edit the size, or reject it and wait.
Rules the swarm can never break
These rules are the reason a human-in-the-loop desk is safer than an autopilot trading bot.
- No order without you. Agents write plans; only you can act on one.
- No plan without a red-team. Cassandra reviews every case before Quill drafts anything.
- No overruling the risk gate. A VETO from Aegis stands even when every other seat agrees.
- No return promises. No agent may state a profit target as a promise or invent a yield.
- No silent disagreement. Objections are kept with the plan so you see the full debate.
- No market without an owner. Stocks, options, crypto, and macro each have a named specialist.
AI trading agents FAQ
Why does Astra use eight AI trading agents instead of one?
One model asked for an opinion tends to agree with the person asking. Splitting the work into eight narrow roles, with a mandatory skeptic and a separate risk gate, makes disagreement part of the process instead of an accident.
Can the agents trade without me?
No. Quill can only write a plan stamped PLAN NOT ORDER, and Aegis can only pass, condition, or veto it. The only role that can act on a plan is you.
What happens when the agents disagree?
Disagreement is recorded, not averaged away. Cassandra’s objections travel with the plan to the Approval Card, and Aegis can veto even when every other seat is in favor.
Which agent should I start with?
Start with the seat that matches how you already think: Nova for research, Pulse for sentiment, Atlas for macro, Vega for options, Orbit for crypto, Quill for planning, Aegis for risk, or Cassandra for a skeptic. The whole desk still reviews every plan.
Ready to meet the desk?
Astra: HITL AI trading personas. Not a broker. No % return claims. You approve every plan. Trading involves risk of loss.