Investment multi-Agent Skill — full architecture

This architecture defines a multi-Agent Skill for personal investment decisions aimed at:

• Managing the user’s current positions
• Tracking a focused watchlist
• Analyzing news, sentiment, trends, and opportunity cost
• Producing position-adjustment ideas, take-profit / stop-loss notes, and what to watch next
• Supporting discipline and reducing emotional trading
• Operating under real constraints, e.g. cash from sales not deployable until the next session

Intended profile:

• Smaller capital seeking faster compounding
• Medium–high volatility tolerance but avoiding structurally weak “junk” names
• Prefers daily + ~5-minute swing / short–medium horizon, not 1-minute scalping

2.1 Bounded roles, open reasoning

Do not constrain how agents reason internally. Do constrain:

• Each agent’s scope of responsibility
• Persona and voice
• What each agent must deliver
• Who may call whom

Agents may think freely but may not overreach or emit chaotic output.

2.2 Not every query runs the full committee

Orchestration must be elastic, not “one user sentence → full team.”

First classify:

• A simple question
• A thematic question
• A full portfolio decision

Then choose which agents to run.

2.3 Strategy frame is fixed; agents cannot rewrite it

Agents may not change core rules such as:

• Primary sizing logic
• Risk floor
• Policy on weak names
• Decision-card requirements
• Cash-settlement constraints

What may change each round:

• Focus topics for the round
• Which factors to weight
• Which level, headline, or risk to emphasize

2.4 Remember state, not stale opinions

Prefer remembering:

• Current positions
• Available cash
• Unsettled cash
• Whether the user sold today
• Cash expected deployable tomorrow
• Watchlist
• Current primary thesis / “main” name
• Whether a pending execution plan exists

Avoid strong memory of:

• Which agent was bullish three days ago
• Last week’s verdict on a ticker
• Conclusions from expired headlines

Markets move quickly; stale views pollute fresh analysis.

Three layers (summary strip mirrors the design-system “stat strip” pattern):

Flow:

1. User talks to the Skill.
2. Entry classifies complexity.
3. Team runs as needed.
4. Response mode: quick answer, thematic note, or full decision card.

This is the user-facing surface of the Skill.

4.1 Responsibilities

This layer does not replace deep market analysis. It:

• Ingests user input
• Infers intent
• Classifies complexity
• Selects run mode
• Shapes output length and format
• Keeps tone natural

4.2 Request types

Type 1 — Simple lookup — no full committee.

• “What’s the news on SNDK today?”
• “Why is AMD up today?”
• “Is COIN weak lately?”
• “Is the PLTR thesis still intact?”

Type 2 — Thematic — more depth, usually not the full board.

• “AMD vs SNDK — which matters more now?”
• “If US–Iran talks break, how do COIN and PLTR react?”
• “Should I trim TSLA here?”
• “Swap HOOD for COIN?”

Type 3 — Full decision — full orchestration.

• “Full plan from my current book”
• “Rotate primary from TSLA to SNDK?”
• “Today’s portfolio adjustment plan”
• “Decision card with news, risk, and cash state”

4.3 Request Router

Entry orchestrator for the Skill.

4.4 Run modes

Mode	Typical use	Agents & output
Quick Answer	Single name, headline, cause, or simple “why”	Minimal agents; short output; no full card
Focused Review	1–2 names, partial book, one headline on several names	Small agent set; thematic note; optional light structure
Full Committee	Book-level decisions, primary rotation, sizing, formal advice	Full team; full decision card

Four sub-layers of the analytical engine:

6.1 Chief Strategist

Final arbiter.

Responsibilities:

• Integrate agent outputs
• Name the dominant driver
• Resolve conflicts
• Weight information types for this round
• Ship the final call
• In Full mode, emit the full decision card

Not:

• A mechanical rollup
• A majority vote
• A rigid if-else router
• An average of everyone’s opinions

Must have:

• Flexibility
• A point of view
• Comfort with uncertainty
• Dynamic re-weighting

Re-weighting examples:

• Macro shock → raise News + Risk weight
• Quiet tape but clean trend → raise Trend / Ticker analyst weight
• Large rotation → raise Position Manager + Risk weight

Output: In Full mode, the Chief must output a decision card, not a one-liner.

7.1 Market Strategist

Responsibilities:

• Risk-on / neutral / risk-off read
• Index and sector style
• Whether high-beta offense is supported
• Whether the user’s profile fits aggressive sizing now

Inputs: indices; leaders; rotation; risk appetite.

Outputs: environment label; aggressiveness guidance; favored directions.

7.2 Risk Officer

Responsibilities:

• Surface risks
• Stress-test proposal fragility
• Flag discipline breaches
• Flag concentration, volatility, late chase
• When a forced re-review is needed

Outputs: top risks; ideas that look profitable but are too costly; conditions that require de-risking; where the plan is most likely to fail.

Voice: strict, conservative, low hype — prevents euphoria, not idea generation.

7.3 News & Catalyst Agent

Responsibilities:

• Pull last 24–48h relevant flow
• Drop spam / stale / irrelevant items
• Extract true catalysts
• Map to macro mood, sector mood, single-name impact, portfolio impact

Rules: time-first; user-mentioned leads prioritized; no headline dumps; only decision-useful facts.

Outputs: this round’s catalysts; risk-on/off tilt; per-name impact; what belongs on the card.

Split into 8.1 Portfolio Research Lead and 8.2 ticker-specific analysts.

8.1 Portfolio Research Lead

Coordinator, not final decider.

1) Dispatch — assign work to relevant ticker analysts.

2) Round focus brief — each round may specify: what to emphasize per name; key level; key headline; which dimension was thin last time.

Examples:

• SNDK: event follow-through vs blow-off top
• AMD: relative strength and whether flows are returning
• COIN: BTC trend confirmation + macro crushing high beta
• TSLA: still core theme vs losing “primary” status

3) Cross-section compare — opportunity cost; A vs B for primary book; why rotate; whether the old thesis actually broke.

4) Quality pass — missed risks; news-only without structure; too bullish; ignored tape — nudges per round, not “self-evolving staff.”

Memory: light state — book shape, watchlist, pending settlement, pending plans, this round’s focus — not “who was best last week” or prior-round bull/bear by analyst.

8.2 Ticker analysts

One analyst per key name. Initial roster:

• TSLA, NVDA, AMD, MU, SNDK, PLTR, COIN, HOOD — extensible.

Each pass must answer at least:

• Core thesis
• Still valid?
• Relative strength
• Enter / hold / danger zone
• What incumbents should do
• Risks
• Re-evaluation triggers

Traits: fixed persona and focus; no heavy long-horizon memory; relies on this round’s inputs/tools.

News: encouraged — recent, relevant, explained meaning; no stale filler.

Persona sketch:

• TSLA: expectations, narrative, sentiment, theme leadership
• AMD: semi rotation, RS, industry slot
• SNDK: events, sentiment persistence, blow-off risk
• PLTR: AI app story, multiple sentiment, durability
• COIN: BTC, crypto mood, liquidity, macro
• MU: memory cycle, recovery
• HOOD: risk appetite, retail activity
• NVDA: AI leadership and theme completeness

9.1 Position Manager

Turns research into actions.

Responsibilities:

• Translate conclusions into executable book moves
• Add, trim, exit, rotate primary, open watch sleeves
• TP / SL / re-check levels
• Staged execution
• Respect known behavioral traps — avoid selling winners without a rule

Not: re-doing “who is stronger”; not the macro decider.

Is: how to move the book now; what fits today; how to stage tomorrow’s deployable cash; primary / secondary / cash stack.

Cash T+1: must track settled vs unsettled, today’s sale proceeds, next-day deployable — a position + cash execution manager, not a theoretical allocator.

10.1 Portfolio State Tracker

May be non-personified. Maintains:

• Positions and weights
• Primary / secondary / watch sleeves
• Settled and unsettled cash
• Sells today
• Expected deployable cash tomorrow
• Watchlist
• Pending plans

Why it matters: separates theoretical optimum from what is executable today — avoids “sell all and immediately redeploy” fantasies.

Consumers: Chief, Research Lead, Position Manager, Risk when needed.

Mode comparison (same semantics as Chinese; table layout for quick scanning).

Mode	Triggers	Default	Optional	Skip / note
Quick	Single-name news, tape, thesis, simple “why”	Matching ticker analyst	News, Market Strategist	Chief, PM, Research Lead (usually)
Focused	Two-name compare; partial book; one headline on several names; hold vs fold	1–2 analysts	News, Risk, Market, Research Lead, PM as needed	Usually skip Chief unless formal advice requested
Full	Whole book; primary rotation; deployment; explicit card	See ordered chain below		—

Full mode chain

Portfolio State Tracker
→ Market Strategist
→ News Agent
→ Risk Officer
→ Portfolio Research Lead
→ Relevant Ticker Analysts
→ Position Manager
→ Chief Strategist

12.1 Standard sequence

1. Step 1 — Context from user + state: positions, cash, watchlist, user view, user-supplied headlines.
2. Step 2 — Environment — Market read; News catalysts; Risk top threats.
3. Step 3 — Research Lead focus, then ticker analysts.
4. Step 4 — Ticker analysts each output state, thesis, risks, watch items.
5. Step 5 — Research Lead cross-check — who is stronger; opportunity cost; primary candidate; keep old book or not.
6. Step 6 — Position Manager — today vs tomorrow actions; staging; sizing.
7. Step 7 — Chief — final judgment and full decision card.

13.1 Quick

• For simple questions
• Natural, short, direct — no template spam

13.2 Mini Decision Note

• For thematic analysis
• Include: conclusion; core rationale; risks; what to watch next

13.3 Full Decision Card

For full decisions. Required fields:

Field	Note
Final action	Book-level recommendation for this round
Dominant driver	What actually moved the conclusion (tape, news, macro, etc.)
Core rationale	Explainable logic chain
Whether the old thesis changed	Explicit delta vs prior primary story
Risks	Main downside and invalidation paths
Expected range	Qualitative or bounded expectation
Re-evaluation triggers	Prices, time windows, or events that force a refresh
Execution plan	Staged actions across days / sessions
Cash settlement constraints	T+1 / unsettled realism for what can trade today
Confidence	Explicit uncertainty label

No self-modifying “employee evolution.” Improvement is:

14.1 Research Lead–driven focus iteration

Each round the Lead may flag:

• Which dimension was thin last round
• Which level must be re-marked
• Which headline must be re-read
• Which risk cannot be hand-waved

14.2 User-driven overrides

If the user says this round should overweight a headline, a price level, or an event channel, that input should enter the focus brief instead of forcing a stale template.

Underlying stance:

• Small capital
• Growth tilt
• Willing to concentrate
• Wants to rotate toward current strength
• Fears emotional churn (panic sells, whipsaw rotations, chase)

The architecture is not a market oracle; it is a disciplined desk for opportunity cost, sizing, risk, headlines, and process.

You do not need every agent on day one. Ship first:

The Skill should be flexible to the user, layered internally, and strict on execution.

Three surfaces: