Multi-Agent Workflow Designer

Wrote docs/architecture/tasks/contract-review-001/agent-workflow-design/multi-agent-workflow-blueprint.md. Short version: you do not need three agents. The job is a bounded pipeline with two model-controlled steps and a deterministic check that makes an invented clause impossible to report. It fits your budget at 48 model calls in the worst case.

Contract Review Workflow Blueprint [WF-CR-1]

Summary

The current researcher, summariser and critic share one context and differ only in prompt, so the critic re-reads the summariser's errors instead of the contract. That is how an invented clause was approved. This blueprint replaces them with four steps: extract, classify, verify, assemble. Only extract and classify call a model. A clause reaches the report only if its text is found on the cited page by a deterministic check. The graph has no cycles, at most 48 model calls, and a 5-minute hard stop that yields a report marked incomplete, never a silent failure.

Terminal Outcomes

The workflow never sends a report. Sending happens only after the lawyer signs off.

Workflow Graph

Worst case: 8 + 40 = 48 model calls, leaving 12 for retries inside the 60-call budget.

Edge Contracts

N1 → N2: JSON array, schema-validated. page must be inside the window; text must be non-empty. A malformed item is dropped and logged, never repaired by a model.

• N2 → N4: every clause_id must exist in N1 output; unknown IDs are rejected at the edge.

N1 → N3: every clause is checked against pdf_text(page) after whitespace normalisation. found_on_page: false removes the clause from the report and lists it under "rejected extractions".

N3 → N4: the report includes a clause only if found_on_page is true. This is the rule that stops last week's failure.

Graph Bounds

• Depth 2 (window → type); no node spawns another node.
• Fan-out: at most 8 windows, at most 40 clause types.
• Cycles: none. No re-review loop.
• Budget: 60 model calls and 5 minutes per contract, enforced by the orchestrator, not by prompts. Reaching either ends the run as incomplete.

Merge and Approval Authority

• N4 is deterministic code; no model merges results.
• No node approves its own output: N3 checks N1, and the lawyer approves the report.
• The "critic" role is removed. A model asked to approve a predecessor that shares its context is not an independent check.

State, Retry and Resume

• Tools are read-only, so a repeated call has no side effect.
• Each window and each clause type is checkpointed when it completes. A resumed run skips completed work.
• One retry per failed model or tool call, only while budget remains. A second failure marks that window or type unprocessed.

Evaluation

• N3 guarantees no reported clause is absent from the contract.
• It does not catch missed clauses. Measuring misses needs contracts labelled by legal-ops (handoff H-3).

Handoffs

Explicit Unknowns

• Which model runs N1 and N2, and its latency: the 5-minute limit is enforced, but not proven to fit 48 calls.
• How pdf_text handles scanned pages without a text layer.
• Clauses that match more than one playbook type.

Traceability

Verification

No validator was supplied, so no command was run. Self-checks against the request:

• call ceiling 48 ≤ 60;
• every edge has a schema and a rule for malformed input;
• no node certifies its own output.

Next steps

1. Legal-ops confirms the three terminal outcomes and the "incomplete" report format.
2. Prompt owner writes the N1 and N2 prompts against the edge schemas (H-1).
3. Build N3 first: it is the cheapest step and closes the invented-clause failure on its own.