◆ Autonomous Worker
When the company reports or the data prints, it pulls the numbers, refreshes the model, and surfaces what changed versus the prior estimates as a diff. It updates the model and hands the diff to the note-drafting agent, which carries the view forward; it does not publish.
Memory
Working The release, the prior model, the recomputed estimates + diff.
Episodic The company's reporting history and estimate accuracy.
Semantic Accounting conventions, the house model structure.
Store Model store + filings warehouse
Orchestration
pipeline MCP
Harness · Managed Agents … event-driven on data releases; sandboxed model execution.
Tools
{ } Filings + data-release feed API ›_ Model execution sandbox Code exec { } Estimates database API
Evals & guardrails
- Extraction accuracy validated against the source filing; figures cite the line item.
- Material estimate changes handed to the note-drafting agent, never published from the refresh itself.
- Model-integrity check before the refresh is committed.
Frontier edge
- ▲Multimodal extraction: reads the filing's tables, footnotes and earnings-deck charts natively, not just the structured XBRL feed.
- ▲Proactive refresh: pre-stages the model the instant the release crosses the wire, so the drafting agent inherits an updated model not a deadline.
- ▲Causal reasoning: explains why an estimate moved (mix vs. one-off vs. guidance) in the diff, not just the delta.
In numbers
1,400
Models refreshed / season
< 4 min
Refresh latency post-release
Handoffs
Hands to → Research Note Drafting Agent