Doer and Architect
Ramp Labs pairs an internal finance agent (the Doer), which executes spreadsheet accounting workflows such as monthly revenue reconciliation and month-end inter-company transfers into ERP, with a second agent (the Architect) that watches screen recordings of accountants and writes the prompts and context files the Doer needs.
- Company
- Ramp
- Approach type
- Agent family
- Work
- Finance ops
- Human involvement
- Unknown
- Invocation
- Background
- Deployment stage
- Deployed
- Evidence strength
- Detailed primary
- Entry reviewed
How it works
Accountant screen recording → reusable Process → completed spreadsheet.
An accountant uploads a screen recording; the Architect extracts the workflow, the required files, and temporal dependencies, and generates a structured prompt[Source 1]
The Doer receives the Process and uploaded files, manipulates the spreadsheets, and returns a completed spreadsheet[Source 1]
Where people stay involved
People review, approve, or guide the agent’s work here.
Reported: The post reports human intervention only after partial completions; the review boundary for completed spreadsheets and ERP entries is not documented.
Supervision level
Not classified
Implementation details
- Model
- Architect uses Gemini 2.5 Pro with a 1M-token context window; Doer uses Claude Sonnet 4.5[Source 1, Source 1]
- Harness
- A two-agent pipeline in which the Architect compiles a reusable Process of steps, required files, and context, and hands it to the Doer to execute; the Doer relies on model intelligence with minimal guide rails[Source 1, Source 1]
- Sandbox
- Not reported
- Tool access
- The Doer uses specialized tools to search, manipulate, and validate spreadsheet operations; workloads combine invoice and data-warehouse data, several tools, Excel workbooks, and ERP journal-entry import[Source 1, Source 1]
- Knowledge
- A Process is a compiled set of steps, required files, and additional information such as temporal dependencies, reusable across runs[Source 1]
- Context management
- Not reported
- Credentials
- Not reported
- Interfaces
- Not reported
Mechanisms
The Architect compiles steps, required files, and context into a Process that later runs reuse[Source 1]
Validation and failure handling
The Doer runs validation checks as it executes each instruction[Source 1]
Reported observations
These results come from the sources. Notes below each result explain what it covers and any reported limits.
Tasks that took 1-2 hours complete in under 10 minutes, a reported 6-12x speedup[Source 1]
- Reported by
- Ramp
- Scope
- Spreadsheet accounting tasks that took 1-2 hours, such as monthly revenue reconciliation and month-end inter-company transfers
The source does not report the denominator of this figure.
Observed in October 2025
About this result
- Category
- Effectiveness
- Evidence type
- Reported measurement
- What it covers
- The named spreadsheet accounting tasks
49.5% accuracy on SpreadsheetBench soft-restriction tasks, tested on about 50 randomly selected tasks with 3 test cases each (OpenAI reported 45.5%)[Source 1]
- Reported by
- Ramp
- Scope
- About 50 randomly selected SpreadsheetBench tasks with 3 test cases each
The source does not report the denominator of this figure.
Observed in October 2025
About this result
- Category
- Effectiveness
- Evidence type
- Reported measurement
- What it covers
- Doer agent accuracy on SpreadsheetBench soft-restriction tasks
32.5% accuracy on SpreadsheetBench hard-restriction tasks in the same test (GPT-4o reported 13.38%)[Source 1]
- Reported by
- Ramp
- Scope
- The same about 50 randomly selected SpreadsheetBench tasks with 3 test cases each
The source does not report the denominator of this figure.
Observed in October 2025
About this result
- Category
- Effectiveness
- Evidence type
- Reported measurement
- What it covers
- Doer agent accuracy on SpreadsheetBench hard-restriction tasks
30-45 minutes of prompt crafting, file uploading, and error resolution per use before the Architect existed[Source 1]
- Reported by
- Ramp
- Scope
- Prompt crafting, file uploading, and error resolution per use of the finance agent
The source does not report the denominator of this figure.
Observed in October 2025
About this result
- Category
- Cost latency
- Evidence type
- Reported measurement
- What it covers
- Manual setup time per finance-agent use before the Architect
Lessons
At Ramp the bottleneck was instruction, not execution; once the Architect wrote the prompts, adding workflows cost near-zero human time instead of scaling with prompt-writing capacity.[Source 1, Source 1]
Ramp Labs used Gemini 2.5 Pro's long-context multimodal ability to extract workflows from video and Claude Sonnet 4.5 for swift, precise, economic execution.[Source 1]
Despite powerful models, the Ramp Labs team spent weeks iterating on system prompts, tool designs, and error handling; the engineering mattered as much as the models.[Source 1]
Catalog interpretation: The post documents unattended prompt generation and asynchronous pickup, but not whether a person reviews the completed spreadsheets or ERP entries; the catalog records the review boundary as unknown.[Source 1, Source 1]