The sidebar contains all navigation items. It collapses to icon-only mode by clicking the chevron button on its right edge to give more space to the main content.

The central area switches between 16+ views depending on which sidebar item you click. Each view is self-contained with its own toolbar, search and content. Only one view is visible at a time.

The Deep Search view includes 8 pre-built query templates: Company Latest News — Funding Round Research — M&A Deal Intelligence — Sector Trends — Executive Profile — Earnings Preview — Macro Event Impact — Regulatory Research.

Before every agent run, MemoryEngine.buildMemoryContext(query) is called. It retrieves relevant memories from all three tiers and prepends them to the agent's system prompt. This means:

The Sandbox view has a template picker: Monte Carlo Simulation — DCF Valuation — Black-Scholes Option Pricing — VaR Calculation — Cap Table Model — Waterfall Simulation — Risk Parity. Each is a fully working Python script you can run immediately and modify.

In your system prompt, directly reference custom field names: "Analyze the company {{company_name}} in the {{sector}} sector at {{stage}} stage targeting a {{check_size}} investment." When users fill these fields before starting a session, they are automatically injected into the prompt context. This eliminates ambiguous questions and makes the first response immediately useful.

In the chat input, type @agent to get a quick-select popover of your recent agents. Select one to immediately start a session without opening the Agent Selector modal. This saves 3–4 clicks for frequently used agents.

Do not use "Use all 119 tools" unless you are building a General Finance agent. Gemini reads all tool declarations before each step — more tools = longer prompts = higher cost and slower responses. A specialized agent with 8–12 precisely relevant tools will outperform a 119-tool agent in both speed and quality for a specific domain.

Deep Search (Gemini Grounding) adds latency and cost. The agent auto-detects when web search is needed based on keywords like "latest", "recent", "current price", year mentions, or ticker patterns. For purely analytical questions (DCF calculation, ratio interpretation), keep Deep Search off to get faster, cheaper responses from the model's training data.

Start every new company research by first asking the KG: "What do we know about [Company]?" If it returns nothing, your agents will build the graph as they work. After 3–4 research sessions on the same company, the KG will have a rich node with facts, relationships to known funds/competitors/people, and historical analysis summaries — making each subsequent session dramatically faster and more accurate.

After running a DCF or valuation analysis, open Time-Travel — find the step where the agent fetched financial data — use What-If mode to change a key assumption (e.g. WACC from 10% to 12%) — re-run. You get a parallel scenario analysis without re-typing the entire query. Save both branches as named scenarios for board presentation.

ReAct agents respond to explicit instructions in the system prompt. Add these high-impact lines to any agent:

These explicit instructions dramatically improve tool utilization and response quality without any code changes.

Create a standard set of team agents (Deal Screener, DD Analyst, Valuation Engine) and promote them to Team Agents with Edit permission for senior analysts and View permission for associates. Every team member runs the same agents — all results appear in the shared Analytics view — the Spawn History provides a searchable audit trail of all multi-agent DD runs across the team.