{ "video": { "id": "5Bn0xro2ol8", "title": "How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, Figma", "duration": 1063, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "5Bn0xro2ol8", "title": "How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, Figma", "one_liner": "A Figma engineer's field report on rolling out coding agents across an engineering org: invest in verification, spend a week writing the plan and let the agent implement it overnight, and put your AI skeptics in charge of the roadmap for making agents safe.", "summary": "Eyal Blum describes Figma's internal (still unfinished) journey adopting coding agents while protecting code quality. He frames adoption as a three-act story — easy 10x wins, then painful failure on bigger problems that destroys trust, then the real skill of guardrails, prompting and context — and notes adoption is uneven across teams that still have to ship together. The concrete practices he offers are: left-shift everything from human review down to deterministic checks and agent review, use TDD-style red/green so the agent fits code to the verification criteria rather than the reverse, write long detailed plans broken into independently-verifiable phases, and mark clearly in PRs, Slack and email which text was written by a human versus generated by AI.", "key_points": [ "Three acts of AI adoption for both individuals and orgs: pick something up and get simple things working 10x faster; apply the same practices to bigger problems and get bad code and bugs, breaking the trust you built; then learn the real skill of guardrails, prompting and context. Teams sit in different acts simultaneously and still have to ship together.", "Three friction points observed: reduced developer agency costs job satisfaction (engineers used to take pride in getting into flow, now they wait on output in a prompt cycle and burn out); the best engineers — the ones holding institutional context in their heads and preventing bad code with 'mental duct tape' — become bottlenecks and are slowest to adopt because they see the problems first-hand; and design docs, Slack messages and emails have become 3-4x as long with 2-3x as many emails while saying the same amount.", "Investing in verification is the highest-value thing you can do in a codebase — every time you left-shift something from a human doing it to an agent verifying it. Playwright plus MCP was named as a specific unlock: agents explore the app instead of humans navigating it.", "When an agent finds something useful, encode it into a deterministic flow — it repeats easily, saves tokens and time, and reserves the LLM for actual reasoning.", "TDD red-green-red-green almost always gives better results, because writing code first and tests after means the agent fits the test to the code rather than fitting the code to the verification criteria.", "A testing-pyramid analogue for agent work: push as much as possible down to deterministic analysis (linting, compiler, unit tests), let agents review against encoded architectural standards in the middle, and leave only functionality and 'is this the right thing to build' at the top for humans.", "Plan-over-prompt restores the craft: it is not uncommon to spend a week writing a detailed plan, making decisions, iterating and sending it to teammates for review, then hand it to an agent to implement. A good plan starts with a bold 'why' / executive summary to prevent agent drift, breaks into parts small enough that you'd want to review the corresponding PR in one sitting (the test: 'I'm going to need to get a cup of coffee before I read this' means it's too big), and puts a validation gate on each phase so later phases don't build on unvalidated assumptions.", "Claimed result: ~20 PRs, none bigger than about 100 lines, from roughly two plans — about six weeks of pre-AI coding work (one week of planning plus a week aligning with three other teams, then an overnight agent run) compressed into one week, a 5x speedup including the review cycle.", "Skeptics' feedback is the roadmap for improving how agents interact with the codebase; put them in charge of making AI safe in the org rather than trying to make them use it, and they come along once the improvements make their own lives better.", "Attention-aware communication: human attention is the scarce resource, so mark what was AI-generated versus hand-written. His team starts every PR description with a short hand-written summary, with the AI description below it, so reviewers know where to be suspicious.", "Meet people where they are — tagging an agent in a Slack thread to close the loop on a small request is one of the most powerful adoption moves, especially framed non-passive-aggressively as 'let's try to see if the agent can get it this time.'" ], "takeaways": [ "Left-shift verification: move checks from human review into linting, compilers and unit tests, then agent review against encoded architectural standards, leaving humans only functionality and product-correctness judgement. Whenever an agent discovers something useful, encode it as a deterministic test rather than re-reasoning it every run.", "Prompt agents in TDD style — set the goal as a failing test first — so the agent fits code to the verification criteria instead of fitting tests to whatever code it wrote.", "Shift your own effort from prompting to planning: write the detailed plan, open with a bold 'why' to stop drift, split it into parts whose PRs you'd review in one sitting, and give each phase an independent validation gate before the next builds on it.", "Recruit your AI skeptics as the owners of the roadmap for making agents safe in your codebase, and take their feedback as the actual list of what to fix.", "Label AI-generated versus human-written text in PR descriptions, Slack and email so readers know where to spend their scarce attention — and never send an unlabelled AI-generated analysis to a senior engineer.", "Lower the barrier by letting people use agents where they already are, e.g. tagging an agent into a Slack thread to close a small loop." ], "topics": [ "coding agents", "ai adoption", "engineering culture", "verification", "tdd", "planning", "code review", "developer experience" ], "tools": [ "Figma", "Playwright", "MCP", "Slack", "Claude Code (cloud agents)" ], "quotes": [ { "text": "investing in verification is probably the highest value thing we can do in our code base", "at": "05:06", "url": "https://www.youtube.com/watch?v=5Bn0xro2ol8&t=306s" }, { "text": "it will almost always give you better results than writing the code and then writing the test afterward because then it will fit the test to the code rather than fit the code to pass the verification criteria", "at": "06:31", "url": "https://www.youtube.com/watch?v=5Bn0xro2ol8&t=391s" }, { "text": "If it's going to be too big for me to want to review in one sitting, it's kind of like the test is I'm going to get need to get a cup of coffee before I read this.", "at": "09:17", "url": "https://www.youtube.com/watch?v=5Bn0xro2ol8&t=557s" }, { "text": "They're skeptic because they're seeing the the way you are lacking validation, where your tools fail. So, and their feedback is basically the road map of how to improve your agent interacting with the code base.", "at": "11:59", "url": "https://www.youtube.com/watch?v=5Bn0xro2ol8&t=719s" }, { "text": "In the age of AI, human attention is a scarce resource.", "at": "12:58", "url": "https://www.youtube.com/watch?v=5Bn0xro2ol8&t=778s" } ], "words": 3549 }, "summary_url": "/#5Bn0xro2ol8", "transcript": { "html": "/transcripts/5Bn0xro2ol8.html", "txt": "/transcripts/5Bn0xro2ol8.txt", "vtt": "/transcripts/5Bn0xro2ol8.vtt" } }