Eyal Blum

Eyal Blum is a software engineer whose work spans mobile infrastructure, developer tooling, and the adoption of coding agents. At Figma, where he spoke about the engineering organization’s ongoing adoption efforts in 2026, he focused on making automated implementation dependable while preserving engineers’ judgment, satisfaction, and attention.

Mobile engineering at Dropbox

• Cross-platform code sharing: Earlier in his career, Blum worked on mobile engineering at Dropbox. His 2019 analysis of cross-platform code sharing examined Dropbox’s move from a shared C++ layer toward native Swift and Kotlin development. Writing an implementation once had promised efficiency, but custom libraries, build tooling, difficult debugging, and the difficulty of finding engineers with the necessary expertise consumed those savings. His argument made developer experience and maintainability part of the architectural calculation: sharing code also meant maintaining the machinery that allowed two platforms to use it.

Figma infrastructure and sustained AI development

In his 2026 talk, Blum identified himself as a software engineer at Figma. Coding agents increased the importance of that infrastructure: generating code faster also increased the need for reliable checks and ways to make engineers’ unwritten knowledge available to the tools.

Blum described using PRDs, design documents, and implementation plans as context carriers for AI sessions. His planning case study describes going from a Slack thread to a shipped project in 48 hours. The documents preserved decisions between sessions, gave colleagues concrete material to review, and supplied acceptance criteria for checking the implementation.

Making coding agents dependable

Blum’s approach to coding-agent adoption connects these engineering practices with the conditions that make people willing to rely on automation. He describes an uneven progression: early successes encourage engineers to attempt larger tasks, failures erode confidence, and useful adoption then depends on developing better context, verification, and workflows. Within one organization, teams at different stages must still work together. He presented Figma’s experience as an ongoing effort, with automation and cloud-agent workflows still being worked out.

• Deterministic verification: He places deterministic verification at the foundation. Compilers, linters, and tests should handle repeatable checks; when an agent discovers a useful verification procedure, engineers should encode it as ordinary software so it can run again without repeated model reasoning. He also favors establishing tests before implementation. Otherwise, an agent can write tests that accommodate its code instead of checking the intended behavior. Agent review can assess explicit architectural criteria, while human reviewers retain responsibility for judging whether the functionality is appropriate and worth building.
• Planning and engineering craft: Detailed planning also gives engineers a place to exercise craft when agents perform more of the implementation. Blum describes spending substantial time making decisions and having teammates review a plan before handing it to an agent. The plan begins with the reason for the work, which helps prevent the agent from drifting toward a different objective. Implementation then proceeds in pieces small enough to review in one sitting, each with its own validation criteria. Checking each phase prevents later changes from accumulating on untested assumptions. He favors workflows that engineers can adapt to their own needs, provided collaborators can understand and work with them.
• Learning from skeptics: His treatment of skeptics follows the same logic. Experienced engineers often hold the institutional knowledge that catches subtle failures. Agents can increase their burden because they recognize pitfalls that other contributors miss, making them bottlenecks in review. Blum treats their objections as a roadmap for missing context, inadequate validation, and unreliable tools. Involving them in deciding what must improve makes adoption depend on changes that actually reduce their workload and earn their confidence.

Human judgment in everyday collaboration

• Clear authorship: Communication requires similar care. Blum’s team began pull-request descriptions with a short human-written explanation of what changed, followed by any generated description. The distinction directs reviewers toward the author’s own judgment and makes the status of the remaining text clear. He describes learning this lesson after sending senior colleagues an insufficiently labeled AI analysis of their review comments. Their objection prompted him to apologize and recognize that he needed to separate his own intent from generated material whose quality he was asking them to assess.
• Familiar interactions: He also favors introducing agents through familiar interactions. Asking an agent to complete a task in a Slack thread and report back there lets colleagues see the result within an existing conversation. A successful experience gives them a practical reason to try it themselves. For Blum, reliable adoption involves improving both the implementation process and the experience of the engineers who must plan, review, and maintain the work.