The Atlas BigLaw / Big Michael — documentation bound to its code
7 documents

Submit a task to final answer

Follow a legal task from the moment it is submitted over REST or MCP, through the workflow's ordered phases, into per-phase DyTopo rounds, and out as a single Opus-synthesised answer.

Architecture

T0  Root Orchestrator (1)
    ↓ issues RoundGoals each phase
T1  Domain Managers (4)       — research / analysis / drafting / compliance
    ↓ DyTopo: Need/Offer matching → directed comm graph
T2  Epistemic agents (26)     — reason within a specific practice area
T2  Conceptual agents (8)     — own a cross-domain legal concept
T2  Writing agents (13)       — produce a specific document type
    ↓ tool_use agentic loop (allowedTools enforcement)
T3  Tool agents (6)           — web_search, doc retrieval, extraction,
                                translation, citation check, e-signing
    +
    32 connector tools        — CourtListener, Westlaw, Everlaw, Trellis,
                                Descrybe, Ironclad, iManage, Definely,
                                DocuSign CLM, Solve Intelligence, Slack,
                                Google Drive, Box, Lawve AI, TopCounsel

Each DyTopo round:

  1. Every agent generates a Need/Offer descriptor (Haiku, ~10 tokens)
  2. Engine cosine-matches Needs → Offers to build a directed comm graph
  3. Jurisdiction filter: agents tagged jurisdictions: ["US"] are excluded from EU/UK/AU tasks
  4. Matched agents receive routed messages from their Need partners
  5. Agents run full agentic loops with routed messages + inter-round memory → Findings
  6. Findings written to intra-round whiteboard
  7. Findings pass through CitationGate → Debate (Opus) → Verification (Haiku ×10)
  8. Low-confidence or challenged Findings go to human gate before final output
  9. Haiku synthesises whiteboard into round digest → written to inter-round memory store