How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, Figma

Highlights

• Figma's internal AI adoption follows a three-act process involving initial 10x gains, subsequent failure on complex tasks, and the development of proper guardrails and contextual prompting.

• Comprehensive planning before coding involves spending up to a week writing detailed implementation plans and breaking them into small, independently validated parts.

• Shift-left verification utilizes the Playwright Model Context Protocol (MCP) to allow agents to independently explore code instead of relying on human navigation.

• Attention-aware communication practices require explicitly marking AI-generated content in PR descriptions, emails, and Slack messages to manage scarce human attention.

• Engaging AI-skeptical senior engineers by placing them in charge of the roadmap for safe agent usage leverages their firsthand identification of tool failures to improve the system.

Timeline

1. 00:00 - 04:40

   Phases of Internal AI Adoption and Emerging Friction Points

   Figma transitioned into an AI-first tool, prompting an examination of how internal engineering teams adopt AI agents. Adoption is uneven across teams, creating organizational friction between AI-forward groups and experimenting teams. Additionally, design docs, Slack messages, and emails have expanded in volume while maintaining the same information density, which complicates quality navigation.

   • Internal adoption follows a three-act curve moving from initial high productivity to failure on complex tasks and finally the acquisition of proper prompting and guardrail skills.
   • Reduced developer agency leads to lower job satisfaction as engineers shift from creative coding into a repetitive prompt-and-wait cycle.
   • Best engineers act as bottlenecks because their unwritten institutional knowledge forces them to carry the mental burden of patching agent failures.
2. 04:40 - 07:31

   Investing in Verification and Shift-Left Testing

   Verification represents the highest-value investment in a codebase. The testing pyramid is shifted downward toward deterministic analysis, including linting, compilers, and unit tests, leaving human review strictly focused on high-level functionality and architectural standards.

   • Investing in verification shifts tasks from human execution to agent-driven validation, such as using Playwright MCP for code exploration.
   • Encoding useful agent discoveries into deterministic flows saves tokens and execution time while reserving LLMs strictly for reasoning tasks.
   • Instructing agents to write code using Test-Driven Development (TDD) red-to-green styles yields superior alignment with verification criteria.
3. 07:31 - 11:45

   Pre-Planning and Context Management for AI Coding

   Planning replaces the traditional craft of writing code with architectural design. A well-structured plan allows a developer to coordinate with other teams, write specifications over a week, and then dispatch implementation to an agent overnight, achieving up to a 5x speedup including the review cycle.

   • Writing detailed implementation plans over the course of a week restores the joy of building and prevents agents from drifting off task.
   • Plans must start with the 'why' in an executive summary and break down into segments small enough to review before needing a cup of coffee.
   • Validating each phase independently prevents errors from compounding across stacked sub-agent tasks.
4. 11:45 - 17:28

   Cultural Shifts, Skeptic Management, and Communication

   Managing the cultural shift requires treating human attention as a scarce resource. Explicitly delineating human-written text from AI-generated text prevents friction during peer reviews. Allowing developers to tag agents directly in Slack threads helps bridge the gap for team members who are not yet fully bought into the workflow.

   • Involving AI-skeptical engineers in tool improvement harnesses their critical feedback to build effective safety guardrails.
   • Attention-aware communication requires marking AI-generated text in PRs and messages to signal appropriate reading scrutiny.
   • Allowing teams to use AI organically within existing workflows like Slack integrations reduces friction and normalizes everyday usage.