Figma's Coding Agents Playbook Without Garbage

Alon Blum says Figma cracked coding agents by making verification, not prompting, the core job, according to AI Engineer.

Figma (NYSE:FIG) is now the browser editor where design, engineering and AI ship together, and Blum spoke as a software engineer on its internal rollout.

Three acts and uneven reality

Blum describes adoption in three acts: quick wins at 10x, then failure on bigger problems as trust breaks, then real skill with guardrails and context.

Inside Figma that adoption is uneven, with some teams fully transformed and others stuck in experiment or skepticism.

Those groups must still ship one product, so coexistence matters more than a single mandated workflow.

Where the friction actually shows

Reduced agency pushes engineers from flow state into a prompt-and-wait loop that hurts satisfaction and accelerates burnout.

The best engineers get hit hardest because they carry unwritten institutional context and become bottlenecks holding back bad agent output.

Communication bloat follows, with design docs growing three to four times longer and email volume doubling or tripling without adding substance.

Why verification beats prompting

Blum calls left-shifting verification the highest value move, letting agents check what humans used to.

When Playwright and MCP arrived, agents could explore the codebase themselves instead of humans clicking through it.

Anything an agent proves useful at should be encoded into a deterministic check to save tokens, time and reasoning.

Writing code test-first in red to green style also works better, because the agent aims at a goal instead of fitting tests to already-written code.

The pyramid that saves human review

Blum maps this to the classic testing pyramid, pushing as much as possible down to deterministic analysis like linting, compiler and unit tests.

Architectural standards that can be codified move to agent review, leaving humans to judge only what humans must: whether this is the right thing to build.

Planning restores craft

Planning versus prompting became Figma's fix for agency loss, replacing live coding joy with the craft of writing a plan.

Teams now spend a week drafting, debating and reviewing a detailed plan, then hand it to agents for implementation.

A good plan starts with a why to prevent drift, breaks into small independently verifiable phases sized to a single reviewable PR, and gates each phase so errors do not compound.

Blum showed a result of about 20 PRs, each 10 to 100 lines and nothing larger, where work that once took a week of coding plus a week of alignment with three teams shipped overnight from the plan.

He counts that as roughly six weeks of coding compressed to one week, or about 5x speedup after review.

Turn skeptics into the roadmap

Blum warns against forcing skeptics to use AI and instead puts them in charge of what needs fixing for safety.

Their complaints map exactly where validation is missing and tools fail, and fixing that list restores their trust.

Attention is the scarce resource

In an AI-saturated org, human attention does not scale, so Blum pushes attention-aware communication.

His team marks what was written by hand versus generated, putting a short human summary at the top of every PR description so readers know where to focus and where to be suspicious.

Blum learned this after sending AI-assisted analysis of PR comments without clear labeling and hearing senior skeptics call it sloppy, a mistake he now fixes with explicit provenance.

Low-friction entry points help too, like tagging an agent in Slack to close the loop in thread, which normalizes use without mandating fancy workflows.

Why this matters beyond Figma

Every engineering org faces the same pressure seen with GitHub Copilot, Cursor, Windsurf and Devin: agents can write fast but quality collapses without verification.

Figma's answer reframes productivity from prompt tricks to system design, deterministic gates and reviewable slice size.

That shift decides who wins: teams that codify standards and plans outrun teams that chase agent autonomy without guardrails.

How Figma compares

StartupHub.ai data shows Figma scores 66/100, behind Canva at 77/100 but ahead of Cord at 56/100, Vizcom at 55/100, Dessn at 51/100 and Paper at 38/100.

StartupHub.ai data also notes Figma's verified early financing as $3.8M Seed in 2013, a useful anchor for a company now navigating an AI-first rebuild while public.

Against those peers the lesson travels: design-heavy platforms compete on workflow reliability, not just generation speed.

Blum is explicit that Figma has not finished this journey, with automation and long-horizon cloud agents still constrained by dependencies.