Second-Order Automation: Multi-Agent System for Finance Workflow Generation

ramp · Finance · 2026 · source Agentic: yes · promotional bias: med · summary-only (no fetched source) · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: finance workflow automation

Ramp's internal finance agent could execute accounting workflows, but each new workflow still required accountants to spend substantial time crafting prompts, preparing files, and debugging instructions.

Constraints:

• The available source page is only an X card, so implementation details are extracted from the ZenML summary rather than the fetched social page.
• Workflow setup previously took 30-45 minutes of prompt crafting.
• Target finance tasks such as revenue reconciliation and fund-transfer booking could take 1-2 hours manually.
• The system had to make automation usable by accountants rather than only engineers or prompt specialists.

Architecture

Ramp used a two-agent architecture: an Architect agent watches a screen recording of a human finance workflow, extracts the steps and external-data references, and compiles a reusable Process artifact; a Doer agent then executes that process against uploaded files using spreadsheet tools and validation checks.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Accountant screen recording"] --> B["Architect agent"]
  B --> C["Gemini 2.5 Pro"]
  B --> D["Reusable Process artifact"]
  D --> E["Doer agent"]
  E --> F["Claude Sonnet 4.5"]
  E --> G["Spreadsheet tools"]

Components

Data flow

An accountant records a workflow; the Architect agent processes the screen recording with Gemini 2.5 Pro, identifies workflow steps, external data, file references, and temporal dependencies, and emits a reusable Process specification; the Doer agent uses Claude Sonnet 4.5 plus spreadsheet tools to execute the Process on future files and return the completed spreadsheet.

Why this architecture

Ramp separated workflow capture from workflow execution because long-context multimodal understanding and structured spreadsheet execution are different jobs; the Architect creates durable instructions while the Doer runs them repeatedly.

Models used: Gemini 2.5 Pro · Claude Sonnet 4.5

Scale

• Prompt setup fell from 30-45 minutes to roughly a five-minute recording.
  “reducing workflow setup from 30-45 minutes to essentially the time it takes to record their screen (approximately five minutes).”summary, stated
• Tasks that previously took 1-2 hours completed in under 10 minutes.
  “tasks that previously took 1-2 hours now complete in under 10 minutes, representing a 6-12x speedup.”summary, stated
• SpreadsheetBench testing used about 50 randomly selected tasks with three test cases each.
  “they conducted testing against SpreadsheetBench using approximately 50 randomly selected tasks with 3 test cases each.”summary, stated

Alternatives considered (roads not taken)

Have accountants manually craft prompts for every run

Rejected because: Manual prompt-writing made adoption scale linearly with human prompt-writing capacity.

Use one model for both video understanding and spreadsheet execution

Rejected because: Ramp used different models for complementary strengths in multimodal context capture and structured execution.

Treat each automation as a one-off prompt

Rejected because: Reusable Process artifacts let future executions reuse the workflow without recreating instructions.

Patterns

multi-agent-orchestration (agent-architecture)

Ramp split automation generation and automation execution across Architect and Doer agents with different model/tool roles.

tool-use-orchestration (agent-execution)

The Doer agent executes spreadsheet workflows through specialized tools and validation checks rather than freeform text alone.

model-routing (model-selection)

Different models are used for Architect and Doer roles based on multimodal context-window strength versus execution precision.

eval-driven-development (evaluation)

Ramp benchmarked the Doer agent on SpreadsheetBench and reported both benchmark accuracy and real-world task speedups.

Tradeoffs

Use an Architect/Doer split.

Chose separate agents for workflow capture and workflow execution over one monolithic finance agent prompt.

Rationale: The split lets each model/toolchain handle the part it is strongest at and produces reusable process artifacts.

Use a screen recording as workflow input.

Chose recording normal accountant work over requiring accountants to write detailed prompts.

Rationale: Recording lowered the barrier for non-technical users to create automations.

Gotchas (what bit them)

Automation setup became its own bottleneck.

Root cause: The initial finance agent still required accountants to write detailed prompts, upload context, and resolve errors.

Resolution: Use an Architect agent to derive reusable Process artifacts from screen recordings.

Production agents still required substantial engineering iteration.

Root cause: Model capability did not eliminate prompt, tool, and error-handling design work.

Resolution: Ramp iterated on system prompts, tool designs, and error handling over weeks.

Doer benchmark accuracy was far from perfect.

Root cause: Spreadsheet tasks still failed frequently on some categories despite model/tool improvements.

Resolution: Use the system as acceleration with human intervention when partial completion occurs.

Evaluation

Ramp reported SpreadsheetBench testing for the Doer agent, real-world task speedups, and qualitative assessment of the Architect agent's setup-time reduction.

Metrics: SpreadsheetBench soft-restriction accuracy · SpreadsheetBench hard-restriction accuracy · workflow setup time · task completion time · speedup versus manual work

Outcomes

• Workflow setup dropped from 30-45 minutes of prompt writing to about a five-minute recording.
• Some 1-2 hour finance tasks completed in under 10 minutes.
• Doer benchmark results exceeded cited baselines but still showed significant failure rates.

The main scale shift is organizational: workflow creation no longer scales only with human prompt-writing capacity.