{ "video": { "id": "RVxym6mmIns", "title": "How Forward Deployed Engineering is done at Cognition — Jia Wu", "duration": 1058, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "RVxym6mmIns", "title": "How Forward Deployed Engineering is done at Cognition — Jia Wu", "one_liner": "Cognition's deployed-engineering lead argues that coding is now mostly solved and only ~20% of the problem, so forward-deployed engineers earn their keep by mapping Devin onto the test/review/deploy/maintain parts of the SDLC and carrying the field's highest-fidelity eval signal back into the roadmap.", "summary": "Jia, a deployed engineering lead at Cognition, frames forward-deployed engineering as maximizing the overlap between what the company builds and what customers actually need — a two-way loop, not just customer delivery. He argues coding itself is a solved problem given enough context engineering, that pointing agents at a codebase with no direction is just 'token maxing', and that the era's KPI has shifted from token usage to measurable delivery outcomes. He backs it with internal numbers (an order of magnitude more PRs over 6 months despite being behind on hiring) and anonymized plus public case studies (Nubank, a Latin American bank, Bilt), and describes the hiring profile and mantras behind the function.", "key_points": [ "Devin launched in 2024 at 13% on SWE-bench, and within a week the reaction was 'this is actually not that useful' — Cognition leaned into the joke with the SF ad campaign and says 'we're actually good now'.", "Cognition's own product surface spans a CLI, an IDE (the speaker joined via the Windsurf acquisition) and the Devin Cloud agent; single-point tools that are just a CLI or just an IDE can't make a whole organization faster.", "Internally over the last 6 months, being behind on hiring, Cognition shipped almost an order of magnitude more good-quality robust PRs by deploying its own agent; enterprise consumption is growing parabolically.", "Coding is 'mostly a solved problem' — writing code faster is only ~20% of the problem; the remaining value is in how you test, review, deploy and maintain code across the enterprise.", "Measuring ROI on agents is explicitly called an unsolved, ambiguous problem — 'the company that will solve this will be a $5 trillion market cap'.", "A session (one agent trace/trajectory) is the unit of measurement; Cognition derives engineering hours generated and how many are productive from it.", "Three anonymized proof points: a 3-month embed delivering ~+150% headcount equivalent, an ~82% reduction in delivery timelines measured on the same agile metrics before and after Devin, and almost double the raw PRs versus single-point tools.", "Public case studies: a Nubank ETL migration staffed by 50 engineers delivered in about a third of the timeline autonomously; a large Latin American bank's tax identification system migration delivered with half the effort; Bilt merging an order of magnitude more PRs, generating the weekly output of over 10 engineers.", "FDE hiring is T-shaped — wide on people/business/process/customer/technology, with deep spikes; they hire product managers (because if the cost of software engineering goes to zero you need to know how to design a product) as well as founders and spiky engineers, since business sense can be taught but technical depth can't be learned on the job." ], "takeaways": [ "Stop optimizing for token usage or for making individual engineers faster; pick the strategic initiatives with the highest business leverage and measure delivery outcomes (timeline compression, PRs merged, productive engineering hours) before and after the agent lands.", "Aim the agent at the unglamorous SDLC surfaces — backlog remediations, code that never ships, tests nobody writes, automatic alert triage — rather than greenfield prompting, and wire it to respond to specific alerts and events so nobody has to trigger it manually.", "Treat every customer deployment as an evaluation set: classify whether each problem is common across the enterprise or unique to one user, and decide which workarounds, hacks and bugs should become product features, so feedback de-risks the roadmap.", "Budget roughly half your day as an FDE for customer calls and half for hands-on keyboard work — the calls are what surface which initiatives are worth automating.", "When hiring FDEs, screen for the two mantras: relentless curiosity about why a problem is worth solving, and being relentlessly tied in to the customer — with everyone treated as go-to-market." ], "topics": [ "forward-deployed-engineering", "agents", "enterprise-adoption", "roi-measurement", "developer-productivity", "legacy-migration", "hiring", "context-engineering" ], "tools": [ "Cognition", "Devin", "Devin Cloud agent", "Windsurf", "Cursor", "Claude Code", "SWE-bench", "Nubank", "Bilt", "COBOL", "JCL" ], "quotes": [ { "text": "But the problem isn't like writing code faster, that's usually only 20% of the problem.", "at": "04:48", "url": "https://www.youtube.com/watch?v=RVxym6mmIns&t=288s" }, { "text": "We have the highest fidelity evaluation set that comes back from our customers, right? We are in the field every single day.", "at": "07:25", "url": "https://www.youtube.com/watch?v=RVxym6mmIns&t=445s" }, { "text": "You can make engineers like 10x faster. That's fine. That's still valuable. But can you make an organization 10x faster, including every single person that might be technical or non-technical uh across the company?", "at": "11:22", "url": "https://www.youtube.com/watch?v=RVxym6mmIns&t=682s" }, { "text": "We leave nothing on the table. We've deployed somebody in Brazil for like 10 months.", "at": "16:17", "url": "https://www.youtube.com/watch?v=RVxym6mmIns&t=977s" } ], "words": 4112 }, "summary_url": "/#RVxym6mmIns", "transcript": { "html": "/transcripts/RVxym6mmIns.html", "txt": "/transcripts/RVxym6mmIns.txt", "vtt": "/transcripts/RVxym6mmIns.vtt" } }