How It Works
How the Coordinator and Agents work together to answer complex questions
Intent Classification
When a message arrives, the Coordinator classifies intent. Simple factual lookups get an immediate response. Anything requiring analysis, multiple sources, or judgment triggers the full multi-agent pipeline.
The coordinator reads the user model (risk profile, positions, preferences) before deciding how to route the question.
Three-Phase Agent Flow
The Coordinator (Claude Opus) runs an orchestration loop. In each phase it spawns agents, waits for their results, reviews them, then decides what to do next. Results flow back to the coordinator between each phase — it's a loop, not a pipeline.
Each agent runs Claude Opus in an isolated OpenCode sandbox. The coordinator loops between phases, reviewing results before spawning the next batch.
Coordinator Tools
The coordinator (Claude Opus) uses these tools to manage the workflow. Deep research belongs to the agents; the two retrieval tools, web_search and fetch_page, are for one-off reads where spawning an agent would be overkill.
Beyond those two retrieval tools the coordinator reads no data source directly. The finalization tools stay on the wire every turn but are refused until the synthesis turn, so the prompt cache survives.
Spawn a new agent by type, or send a follow-up message to a running one
Stop a running agent that is no longer needed
Show or clear interim synthesis (markdown, or named-key perspectives) while agents run
Create or update a visualization in the user's panel — chart, table, forecast, knowledge graph, etc.
Remove duplicate or withdrawn artifacts from the thread's panel
Record, update, or withdraw a falsifiable forecast with probability and resolution criteria
Record 3 follow-up question chips shown under the final synthesis
Exa web search — titles, URLs, and content excerpts, with category, domain, and date filters
Scrape a specific URL directly via Firecrawl (bypasses agent spawn for one-off reads)
Register one clarifying question with 2-5 options, rendered as a form under the next turn. Interactive threads only.
Run the calibrated EF-32B forecasting model over compressed agent evidence. Registered only when the EF model is enabled for the run.
Trade Terminal Tools
The Trade Terminal variant swaps the default toolset for a structured trade-card workflow. The coordinator can read market data directly, run web search, and progressively fill in the user's Trade Card (thesis → instrument → sizing → risk → finalize). Stage 1 can spawn the full Axion research and analyst set to pressure-test the idea; Stage 2 narrows to the four trade specialists ( tweet-analyst, filings-analyst, earnings-analyst, options-flow-analyst). Artifacts those agents emit are suppressed in both stages so the Trade Card remains the single surface.
Trade Terminal replaces emit_artifact / emit_forecast / suggest_follow_ups with the Stage-1 thesis and Stage-2 structuring tools, and adds direct market-data and options-chain access on top of the shared web_search and fetch_page.
Patch Stage-1 thesis fields on the bound Trade Card
Initialize Stage-2 structure with a single trade instrument
Set Stage-2 sizing for the leg, compute tier / base gross / final gross
Set the delta / beta neutrality target for the structured leg
Set stop, invalidation statement, profit targets, and max hold days
Validate Stage-2 structure and mark trade card status = ready
Current spot and daily OHLC history with ATR computed from OHLC (Massive)
Options chain snapshot for an underlying ticker (Massive)
Exa web search — titles, URLs, and content excerpts with category/domain/date filters
Firecrawl scrape of a specific URL
Stage 1 spawns the full Axion research and analyst set plus the four trade specialists; Stage 2 narrows to tweet-analyst, filings-analyst, earnings-analyst, and options-flow-analyst. Artifacts from spawned agents are suppressed in both stages.
Stop a spawned specialist that is no longer needed
Register one clarifying question for the user. Interactive threads only.
Agent Sandbox Tools
Every agent runs inside an OpenCode sandbox. It reaches the outside world through two surfaces, both provisioned at session start — MCP servers called as native tool_use blocks, and sandbox skills (Python SDKs and HTTP proxies) invoked by writing and executing code. Both are scoped to the sandbox, both are proxied through the backend so credentials never touch the container.
MCP Servers (native tool_use)
Sidecar tools for the sandbox: report_progress, emit_artifact, delete_artifact
One server fronting every retrieval provider: research_exa_search, research_exa_answer, research_exa_deep_research, research_firecrawl_scrape_page, research_youtube_search, research_youtube_transcribe_video
search_tweets, get_user_tweets, get_tweet, get_tweet_thread, get_user, get_user_profile
search_markets — Polymarket + Kalshi odds, prices, and order book depth
Part of the research server. A Claude Sonnet sub-agent searches the internal MediaWiki and returns the answer as markdown.
axion and research are provisioned by the backend at session start; twitter and prediction-markets are external servers named in config. Only axion is fatal to lose — an agent that cannot reach it cannot emit an artifact.
Sandbox Skills (Python SDKs & HTTP proxies)
OHLCV, options chains with greeks & IV, technicals, forex, crypto, short interest, treasury yields, dividends, splits — Python SDK
Mapping, Query, Section Extractor (10-K/10-Q), XBRL-to-JSON, Insider Trading, 13F/13D/G, Form 4, Exec Comp, N-PORT, Form D, S-1/424B4
Options flow alerts, dark pool prints, GEX / gamma exposure, market tide, congressional + insider trades, short selling, crypto whales
800K+ Federal Reserve macro time series (rates, inflation, labor, GDP) + CPI / NFP / FOMC release calendars
Forward EPS & revenue estimates, earnings calendar, analyst consensus, insider sentiment (MSPR), social sentiment, lobbying, gov contracts
Analyst upgrades / downgrades / price target changes + analyst track-record accuracy (via the Massive SDK)
Weekly COT reports — leveraged fund and asset manager positioning in equity indices, rates, FX, VIX, energy, metals
US federal court docket search, PACER filing text, and full case records for litigation / securities fraud exposure
220M+ papers, citation graphs with intent labels, semantic recommendations, author h-index (for peer-reviewed evidence & diligence)
Daily curated ML/AI feed ranked by upvote velocity + semantic search across HF-indexed papers and arXiv IDs
Federal grants, contracts, and SBIR awards — who is receiving US government money, by company, agency, or topic
NIH grant awards for biotech and medtech diligence — SBIR phase detection and topic-based deal sourcing
NSF grant awards for deeptech and AI/ML diligence — SBIR sourcing and principal-investigator background
DeFi TVL, protocol detail, stablecoin supply, yield pools, perpetuals open interest, protocol fees and revenue
Monthly prices for 71 commodities — Brent / Dubai / WTI crude, LNG, coal, metals, agricultural benchmarks
Daily port traffic and chokepoint transits across 2,065 ports and 28 chokepoints for supply-chain and shipping questions
77 FAO domains covering 245 countries back to 1961 — production, trade, prices, food balance, food security
Sub-national retail food prices for staples (wheat, rice, bread, beans, oils, sugar) in food-insecure markets
Search and read Substack publications and posts for independent analysis and primary commentary
System-prompt source hierarchy: primary regulatory + market data (SEC, Massive, FRED, CFTC) > derived / alt signals (Benzinga, Unusual Whales, Finnhub) > news & web.