# How Forward Deployed Engineering is done at Cognition — Jia Wu

Source: [AI Engineer](https://www.youtube.com/watch?v=RVxym6mmIns)  
Feed7 permalink: https://feed7.dev/p/how-forward-deployed-engineering-is-done-at-cognition-jia-wu-06h8ybj  
Published: 2026-07-28T20:00:06.000Z  
Trust: Source Linked (source_linked)

## Why Included

Cognition measures coding-agent deployments by delivery outcomes, not sessions or tokens: engineering capacity, shorter timelines, and accepted PRs tied to customer work.

## Source Summary

Cognition says a three-month embedded deployment produced capacity comparable to **150% additional headcount** and cut delivery timelines by about **82%**. Another cited customer reportedly merged roughly **10× more** work per subscriber.

## Practical Implication

Builders should define business-facing measures before scaling agent usage: accepted changes, cycle time, shipped projects, and maintenance outcomes. Map automations to high-leverage work, then use deployment traces as supporting evidence rather than the goal.

## Agent-Ready Context

Cognition says a three-month embedded deployment produced capacity comparable to **150% additional headcount** and cut delivery timelines by about **82%**. Another cited customer reportedly merged roughly **10× more** work per subscriber.

Builders should define business-facing measures before scaling agent usage: accepted changes, cycle time, shipped projects, and maintenance outcomes. Map automations to high-leverage work, then use deployment traces as supporting evidence rather than the goal.

These are company-presented case studies without baselines, calculation details, or independent validation. Engineering-hour estimates and headcount equivalents can still conceal low-value activity unless paired with accepted, maintained output.

## Connected Context

Feed7 judgment across 262 accumulated Signals:

This supplies unusually large, business-facing estimates for the value of embedded coding agents and sharpens adoption evaluation around accepted, maintained delivery rather than activity. It also narrows how confidently those gains can be generalized: the figures are vendor-presented, lack calculation details and baselines, and therefore establish a measurement direction more strongly than a transferable benchmark.

- [The Dirty Secret of Forward Deployed Engineering — Natalie Meurer, Sierra](https://feed7.dev/p/the-dirty-secret-of-forward-deployed-engineering-natalie-meurer-sierra-17crz97) — Sierra’s outcome-owned view of forward deployment provides the organizational premise; Cognition makes that premise measurable through accepted changes, cycle time, and shipped work.
- [ReviewDebt: a practical framework for scoring every pull request — Sachin Gupta, Ebay](https://feed7.dev/p/reviewdebt-a-practical-framework-for-scoring-every-pull-request-sachin-g-0iyjtyk) — ReviewDebt adds a necessary countermeasure to Cognition’s throughput metrics by testing whether increased output is creating verification burden faster than reviewers can absorb it.
- [ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration](https://feed7.dev/p/scarfbench-1u8lniy) — ScarfBench’s low behavioral success and false build claims caution against treating Cognition’s customer case studies as evidence of broad coding-agent capability without reproducible task-level validation.

## Context Map

- Layer: benchmark
- Domains: coding
- Topics: agent-evals, coding-agents, adoption

## Uncertainty

- These are company-presented case studies without baselines, calculation details, or independent validation. Engineering-hour estimates and headcount equivalents can still conceal low-value activity unless paired with accepted, maintained output.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.