Institution Mode

🟡 Intermediate — long-running analysis (≤25 LLM calls, 60–600 seconds per run). Read-only by design. ⚠️ Risk Disclaimer: institution mode produces opinions, not orders. A "Buy" rating from the Portfolio Manager is still subject to your own sizing, your emergency stop, and the same two-step confirm on every fund-moving call.

Institution Mode convenes a multi-agent investment committee inside Minara. Inspired by the TradingAgents research direction, it runs a structured 6-phase pipeline against any ticker (or basket of tickers) and emits a 5-tier rating with a probability-weighted scenario tree, structured price targets, and a permanent audit log.

The whole pipeline is read-only. It never moves funds. Acting on the recommendation is a separate, explicit step that goes through the same preview-confirm-execute flow as a manual order.

What Institution Mode is for

Institution Mode is the right surface when:

• You are about to make a high-conviction directional decision on a single asset and want adversarial scrutiny before committing capital.
• You need to compare candidates ("BTC vs ETH vs SOL — which now?") or build a cross-asset allocation ("$100k across BTC, gold, the S&P 500, and stablecoins").
• You want a board-style audit trail for a decision: every analyst output, every debate turn, every methodology cited, persisted to SQLite under a single run_id you can reflect on later.
• The user explicitly asks for "institution mode", "机构模式", "multi-agent", "多 agent", a "boardroom" call, or an "investment committee" review.

It is not the right surface for:

• Quick price or balance lookups (use get_price / minara_account).
• Light sentiment checks (use get_fear_greed, social analysts directly).
• Anything where 60+ seconds of latency is unacceptable.

What you get on every run

Every decision-producing role uses forced structured tool-use with a Zod schema, so outputs are deterministic in shape and directly comparable across runs. The full report is saved as report://{report_id} and surfaced inline in the chat. The reflection ladder (1d / 7d / 30d / 90d / 180d / 365d) revisits each run on the schedule above, scoring the call against realised price action and feeding lessons back into the methodology store.

Coverage across asset classes

The pipeline routes each ticker through classifyAsset() (apps/agent/src/learning/methodology-store.ts) to pick the right analyst tool whitelist and methodology bucket for that asset class.

Example prompts

/institution BTC
/institution NVDA --debate 3 --risk 3
/institution BTC ETH SOL
/institution gold SPY USDC --debate 5
/institution 分析下苹果
分析下机构模式 BTC 多 agent 视角

The CLI flags --debate N and --risk N clamp the per-debate round count to 1..5. Without flags, the run uses the INSTITUTION_MAX_DEBATE_ROUNDS and INSTITUTION_MAX_RISK_ROUNDS defaults.

Example output — single-ticker

Prompt: /institution BTC

Agent response (abbreviated):

Portfolio Manager — Buy (confidence 0.72)

BTC is technically constructive (EMA stack bullish on the 4h, RSI mid-range), on-chain is constructive (LTH supply ratio rising, exchange balance flat-to-down), funding is neutral, and the 30d realised vol leaves room for a directional move into the next FOMC.

Scenarios (probability-weighted)

• Base 55% — grind to $122k over 6 weeks; defended at $108k.
• Bull 25% — squeeze through $122k on flow + ETF resumption; target $138k.
• Bear 20% — macro shock → flush to $96k; thesis invalid below $93k.

Trader proposal: Buy spot, size 1.5R against a $93k stop, target $122k base / $138k stretch. Horizon 6–10 weeks.

Risk debate: Conservative warns on funding-flip risk near $122k; Aggressive flags ETF flow inflection; Neutral lands on the base-case sizing above.

Full report: Institution report.

⚠ This is analysis, not an order. To act, ask me to buy 0.5 BTC or long BTC 5x size $200 and I will run the standard preview before any funds move.

Example output — multi-ticker allocator

Prompt: /institution BTC ETH SOL

Agent response (abbreviated):

Allocator — mode: hybrid

Across the three majors, the committee's per-ticker calls are:

Ticker	PM rating	Confidence	Allocator weight
BTC	Overweight	0.71	50%
ETH	Hold	0.55	30%
SOL	Buy	0.68	20%

BTC anchors as the highest-confidence majority hold; SOL gets a tactical 20% on the strongest momentum read; ETH stays on Hold until L2 sequencer fee compression clears.

Acting on this is still a separate step — ask me to swap or buy each leg and I will preview every order.

Cost & latency

A typical run costs ≈25 LLM calls and 60–600 seconds, dominated by Phase 1 parallel analysts and the two debate phases. Tune with:

• INSTITUTION_WALL_CLOCK_TIMEOUT_MS — global budget (default 1200s).
• INSTITUTION_MAX_DEBATE_ROUNDS / _MAX_RISK_ROUNDS — debate depth.

Every role in the pipeline runs on the operator's selected agent model — switch the whole session to a stronger model if you want stronger synthesis (no per-role overrides).

Full env var reference: Environment Variables.

Safety posture

• Read-only. The minara_institution_analyze tool is registered as READ_ONLY. No analyst, debater, or PM can call a fund-moving tool.
• Acting requires a separate prompt. Once you have the recommendation, place the order yourself (swap_tokens, open_perps_position, minara_ss_deploy). Each of those enforces its own two-step preview-confirm gate.
• Audit log. Every run_id writes phase-by-phase outputs, reflections, and methodology citations to SQLite. Inspect via /institution-history in the REPL or minara institution history from the shell.

Skills and tools involved

Where to go next:

• Run your first institution analysis
• How runs feed the learning system
• Institution env vars
• Underlying skill: minara.institution