← back
How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, Figma
Takeaway
Scale coding-agent adoption by investing in verifiable plans and shared engineering knowledge while addressing the human costs of the transition.
Summary
- Figma sees uneven adoption across an initial productivity boost, disillusionment on larger tasks, and eventual skill in guiding agents with context and guardrails.
- Institutional knowledge concentrated in senior engineers makes them review bottlenecks; reduced agency and longer AI-generated communications also create organizational friction.
- Shift verification toward deterministic linting, compilers, and tests, then agent checks, reserving humans for product judgment; test-first development helps avoid tests that merely mirror generated code.
- Detailed plans restore engineering agency: state the purpose clearly, divide implementation into reviewable pieces, and give each phase independent acceptance criteria.
coding-agentsengineering-cultureverification
Original description
The engineers slowest to adopt agents at Figma are the best ones. Eyal Blum's explanation is that they are the people holding the codebase together with mental duct tape, carrying the institutional context nobody ever wrote down, and so they see every failure mode first and trust it least. His answer is not to talk them into using AI. It is to hand them the roadmap for making agents safe in the codebase, because their complaints are an ordered list of where verification is missing, and they come along once the fixes start making their own work easier. He is equally direct about the costs nobody puts on slides. Developer agency drops, and engineers who took real pleasure in writing code find themselves waiting on output instead. Documents and Slack messages have gotten three to four times longer while saying the same thing. His team's fix for the last one is a convention rather than a tool: a pull request description opens with a line a human actually typed, then the generated text below it, so readers know where to spend scarce attention. That habit came from getting it wrong, sending a senior skeptic an unmarked AI analysis and being told it read as slop. Plans get a week; the agent gets the night. Speaker info: / eyalg Timestamps: 0:00 - The three acts of AI adoption 2:28 - Uneven adoption, and teams that must coexist 3:35 - Why the best engineers adopt last 4:45 - Investing in verification first 5:55 - Encoding what works into deterministic flows 7:04 - A testing pyramid for agent review 8:12 - Planning instead of prompting 9:21 - What makes a plan an agent can follow 11:45 - Handing skeptics the roadmap 12:55 - Attention aware communication 15:16 - Sending unmarked AI output, and what it cost 16:29 - Meeting people where they already work