◆ Supervised Orchestrator
Ingests the thousands of documents in a VDR … contracts, financials, cap tables, litigation … and produces a structured findings log: change-of-control clauses, contingent liabilities, customer concentration, risk items. Reads the full room in a single sweep, with every finding cited to a document.
Memory
Working The current document and the open findings checklist.
Episodic Findings already logged on this deal (avoid duplicates / link related).
Semantic Diligence taxonomy and red-flag clause library by deal type.
Procedural Extraction playbooks per document type (contracts, leases, accounts).
Store File-based memory tool + findings knowledge graph
Orchestration
orchestrator-worker MCPA2A
Harness · Managed Agents … orchestrator spawning parallel reader sub-agents by document category with fresh context each; findings merged into a NOTES log.
Tools
{ } Virtual data room API { } Document OCR + parsing API ⌕ Clause / red-flag retrieval Retrieval { } Findings tracker API
Evals & guardrails
- Citation discipline … every finding links to the source document and page.
- Agent-as-judge sampling of findings for false positives / missed red flags.
- Coverage check: flags un-reviewed document categories before sign-off.
Offline reflection
Learns which clause patterns the deal-committee agent escalated vs. ignored, sharpening its red-flag library per sector and deal type.
Frontier edge
- ▲Confidential compute: reads the VDR inside a secure enclave (TEE) so target PII and MNPI never sit in clear text outside the room.
- ▲Multimodal extraction: reads scanned contracts, cap-table images and signed PDFs natively, not just OCR'd text.
- ▲Long-horizon autonomy: chains parallel readers across a 4,000-document room in one supervised multi-hour run on the METR curve.
A sample run
Trigger Buy-side diligence opens on a 4,000-document data room with a tight LOI deadline.
- 1Index and categorise the room; spawn parallel readers per category.
- 2Extract change-of-control, indemnity and termination clauses from contracts.
- 3Flag customer concentration and contingent liabilities from the accounts.
- 4Merge into a cited findings log ranked by deal impact.
Output A structured diligence findings report with top risks surfaced and cited, committed to the deal-committee agent; three high-impact items escalated to an oversight agent.
In numbers
thousands
Documents reviewed / deal
median 6 h on a 4,000-doc room
Diligence sweep latency
Handoffs
Hands to → Pitchbook Builder
Across ⇢ Risk → credit/legal review on flagged liabilities⇢ Compliance → conflicts and MNPI controls