Sign InOpen Brain
AI EngineerVideoSource Linked

How Forward Deployed Engineering is done at Ramp — Leo Mehr

Ramp pairs strict request scoping with agents across the delivery pipeline. Its intake agent reportedly saves about 20% of scoping time, while humans retain final judgment.

AI Engineer · Jul 28, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Ramp’s FDE organization grew from **2 to about 30 engineers**. A Notion-based intake agent now questions requesters until it can draft a spec, cutting reply latency from hours or days to seconds and reportedly saving about **20% of scoping time**.

Practical Implication

Builders should validate urgency, users, workarounds, platforms, and reuse before generating code. Then automate the pipeline in stages—from context gathering and scoping through specs and implementation—with evals, rubrics, and human review at each boundary.

Agent-Ready Context
Ramp’s FDE organization grew from **2 to about 30 engineers**. A Notion-based intake agent now questions requesters until it can draft a spec, cutting reply latency from hours or days to seconds and reportedly saving about **20% of scoping time**.

Builders should validate urgency, users, workarounds, platforms, and reuse before generating code. Then automate the pipeline in stages—from context gathering and scoping through specs and implementation—with evals, rubrics, and human review at each boundary.

Ramp says the middle of that pipeline remains difficult, especially transferring tacit product knowledge into agents. The 20% figure is an estimate, and faster intake can amplify poorly scoped work if judgment is removed.
Connected Context · Feed7 Judgment

This turns process-first agent engineering into a staged operating model: interrogate demand, formalize a spec, then automate implementation behind evaluated human-review boundaries. The reported intake gain confirms value at the context-gathering edge, while the unresolved transfer of tacit product knowledge narrows expectations for end-to-end automation and identifies the middle of the workflow as the real bottleneck.

How Forward Deployed Engineering is done at Kepler — Vinoo GaneshKepler’s discovery loop supplies the field-observation discipline that Ramp’s intake questions attempt to encode before a request becomes implementation work.Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayerHumanLayer reinforces Ramp’s decision to invest in upfront alignment and preserve human review, especially because faster intake can otherwise accelerate poorly scoped work.The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant ContextThe context-sensitivity findings imply that Ramp’s staged evals should test specs and decisions under irrelevant or noisy request context, not rely only on aggregate pipeline scores.State of Data — Sean Cai, Independent / State of DataThe emphasis on live workflow trajectories complements Ramp’s staged pipeline by suggesting that each handoff, recovery, and outcome should be evaluated in operation rather than only through saved answers.
Context Map
agentcoding#harness-engineering#context-engineering#agent-evals
Uncertainty
Ramp says the middle of that pipeline remains difficult, especially transferring tacit product knowledge into agents. The 20% figure is an estimate, and faster intake can amplify poorly scoped work if judgment is removed.