Source: https://internal-agents.com/agents/ramp-finance-agents # Ramp — 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](https://internal-agents.com/organizations/ramp) - Collection: Agents - Approach type: Agent family - Deployment stage: Deployed - Autonomy: Unknown - Evidence strength: Detailed primary - Status: Internal - First reported year: 2025 - Work: Finance ops - Invocation: Background - Entry reviewed: 2026-10-08 ## Purpose ### Summary 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. Fact · Reported · Not assessed confidence · `ramp-finance-agents--summary` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 16-24 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 41-56 Representative workflow: Accountant screen recording → reusable Process → completed spreadsheet. ## How it works ### Architect watches a screen recording An accountant uploads a screen recording; the Architect extracts the workflow, the required files, and temporal dependencies, and generates a structured prompt Fact · Reported · Not assessed confidence · `ramp-finance-agents--primitives-architect-watches-recording` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 70-74 ### Doer executes the Process The Doer receives the Process and uploaded files, manipulates the spreadsheets, and returns a completed spreadsheet Fact · Reported · Not assessed confidence · `ramp-finance-agents--primitives-doer-executes-process` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 82-85 ## Where people stay involved - **Accountant screen recording → reusable Process → completed spreadsheet** — Unknown · Level unknown **Reported:** The post reports human intervention only after partial completions; the review boundary for completed spreadsheets and ERP entries is not documented. ### Supervision evidence #### Operating model assessment Unclassified for Accountant screen recording → reusable Process → completed spreadsheet; human attention boundary: unknown. Inference · Catalog judgment · Unverified confidence · `ramp-finance-agents--operating-models-0` Confidence reason: The post reports that the Architect generates prompts without human intervention and that accountants return to a ready-to-use process, and it names human intervention only after partial completions; it does not state whether a person reviews completed work in a normal run. Qualifications: - Observation date: 2025-10 Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 93 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 102 ## Implementation details ### 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 Fact · Reported · Not assessed confidence · `ramp-finance-agents--architecture-harness` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 45-56 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 62-64 ### Model Architect uses Gemini 2.5 Pro with a 1M-token context window; Doer uses Claude Sonnet 4.5 Fact · Reported · Not assessed confidence · `ramp-finance-agents--architecture-model` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 71 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 83 ### 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 Fact · Reported · Not assessed confidence · `ramp-finance-agents--architecture-tool-access` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 54 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 32-33 ### Knowledge A Process is a compiled set of steps, required files, and additional information such as temporal dependencies, reusable across runs Fact · Reported · Not assessed confidence · `ramp-finance-agents--architecture-knowledge` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 64 ### Compile a reusable Process The Architect compiles steps, required files, and context into a Process that later runs reuse Fact · Reported · Not assessed confidence · `ramp-finance-agents--primitives-compile-reusable-process` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 64 ### Implementation coverage - **Model:** reported - **Harness:** reported - **Sandbox:** unreported - **Tool access:** reported - **Knowledge:** reported - **Context management:** unreported - **Credentials:** unreported - **Interfaces:** unreported ## Validation and failure handling ### Validation checks per instruction The Doer runs validation checks as it executes each instruction Fact · Reported · Not assessed confidence · `ramp-finance-agents--primitives-validation-checks-per-instruction` Confidence reason: No explicit confidence assessment is recorded for this claim. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 84 ## Reported observations Observation: Effectiveness · Reported measurement · The named spreadsheet accounting tasks ### Headline claim Tasks that took 1-2 hours complete in under 10 minutes, a reported 6-12x speedup Metric · Reported · Medium confidence · `ramp-finance-agents--headline-metric` Confidence reason: Ramp Labs reports the comparison directly for its named spreadsheet tasks, but states no measurement method, sample, or time period. Qualifications: - The source does not report the denominator of this figure. - Reported by: Ramp - Scope: Spreadsheet accounting tasks that took 1-2 hours, such as monthly revenue reconciliation and month-end inter-company transfers - Observation date: 2025-10 Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 102 Observation: Effectiveness · Reported measurement · Doer agent accuracy on SpreadsheetBench soft-restriction tasks ### Key observation 49.5% accuracy on SpreadsheetBench soft-restriction tasks, tested on about 50 randomly selected tasks with 3 test cases each (OpenAI reported 45.5%) Metric · Reported · Medium confidence · `ramp-finance-agents--key-metrics-spreadsheetbench-soft-restriction` Confidence reason: A self-run benchmark with a stated task count and comparison figures; the evaluation setup is reported only by Ramp Labs. Qualifications: - The source does not report the denominator of this figure. - Reported by: Ramp - Scope: About 50 randomly selected SpreadsheetBench tasks with 3 test cases each - Observation date: 2025-10 Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 97-99 Observation: Effectiveness · Reported measurement · Doer agent accuracy on SpreadsheetBench hard-restriction tasks ### Key observation 32.5% accuracy on SpreadsheetBench hard-restriction tasks in the same test (GPT-4o reported 13.38%) Metric · Reported · Medium confidence · `ramp-finance-agents--key-metrics-spreadsheetbench-hard-restriction` Confidence reason: A self-run benchmark with a stated task count and comparison figures; the evaluation setup is reported only by Ramp Labs. Qualifications: - The source does not report the denominator of this figure. - Reported by: Ramp - Scope: The same about 50 randomly selected SpreadsheetBench tasks with 3 test cases each - Observation date: 2025-10 Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 100 Observation: Cost latency · Reported measurement · Manual setup time per finance-agent use before the Architect ### Key observation 30-45 minutes of prompt crafting, file uploading, and error resolution per use before the Architect existed Metric · Reported · Medium confidence · `ramp-finance-agents--key-metrics-prompt-crafting-before-architect` Confidence reason: The post states the setup time directly as the reported cost of each use before the Architect existed, with no measurement method. Qualifications: - The source does not report the denominator of this figure. - Reported by: Ramp - Scope: Prompt crafting, file uploading, and error resolution per use of the finance agent - Observation date: 2025-10 Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 16 ## Lessons ### Lesson 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. Fact · Reported · High confidence · `ramp-finance-agents--lessons-learned-bottleneck-instruction-not-execution` Confidence reason: The post states the bottleneck conclusion and the scaling contrast between manual prompt writing and Architect-generated workflows. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 37 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 110 ### Lesson 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. Fact · Reported · High confidence · `ramp-finance-agents--lessons-learned-complementary-model-strengths` Confidence reason: The post names both models and the strength each contributed to its stage of the pipeline. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 111 ### Lesson 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. Fact · Reported · High confidence · `ramp-finance-agents--lessons-learned-production-agents-still-hard` Confidence reason: The post reports the weeks of iteration on prompts, tools, and error handling directly. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 112 ### Lesson 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. Inference · Catalog judgment · Medium confidence · `ramp-finance-agents--lessons-learned-async-intake-not-autonomy` Confidence reason: The inference follows from what the post leaves unstated; the unattended generation and the partial-completion interventions are reported, the review boundary for completed work is not. Evidence: - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 20 - Supports · [1] [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) · [Preserved copy](https://github.com/steel-experiments/internal-agents-map/blob/main/archive/sources/ramp-finance-agents-source-1/content.md) · Preserved content.md, lines 102 ## Question coverage and scope - **purpose:** Reported - **workflow:** Reported - **human involvement:** Reported — The post reports human intervention only after partial completions; the review boundary for completed spreadsheets and ERP entries is not documented. - **implementation:** Reported - **validation:** Reported - **observations:** Reported - **lessons:** Reported ## Sources 1. [We built an agent to prompt our internal finance agent](https://ramplabs.substack.com/p/we-built-an-agent-to-prompt-our-internal) - Engineering blog · First party · Evidence - Original URL: - Publisher: Ramp Labs · Published: 2025-10-01 · Accessed: 2026-10-08 · Last verified: 2026-10-08 - Preserved copy in the repository: