Vercel CEO reveals his Agentic Engineering Workflow

Technological concepts & key product/strategy themes

1) Security debt and “security debt” as a scaling constraint

• The speaker frames adoption of new software/AI as increasingly hard due to:
  • rising attack surface
  • “latent vulnerabilities” in human-written code
• Proposed direction:
  • Move more of the development lifecycle into sandboxes
  • Use more automated “secure-by-construction” pipelines to reduce the feasibility of supply-chain attacks
• Warning:
  • Attackers can leverage strong models to scale vulnerability discovery, so proactive defense matters.

2) Agentic engineering workflow: leverage moves “one abstraction higher”

• The workflow shifts from writing small code snippets to building and operating:
  • “software factories”
  • fleets of agents
  • orchestration across models / tools / sandboxes
• The challenge becomes less about agent capability and more about maintaining human control over:
  • cybersecurity risk
  • how agents operate on tasks
  • evaluation/verification so the system doesn’t generate only fast, low-quality output.

3) Advantage in an “anyone can build anything” world = taste + evaluation + iteration velocity

Competitive edge isn’t just speed. It comes from:

• taste and human creative direction
• timely exploitation of newly released models (the “latent space” exploration effect)
• evaluation/benchmarks to choose the right model/tool for a job
• “personal evals” rather than “vibes-from-Twitter” prompting
• Measurement focus:
  • iteration speed (how quickly things ship)
  • iteration velocity (speed toward the right goal/direction, not endless “PRs in oblivion”).

4) Customer-signal-driven autonomy rubric (“grade” agenticness)

Agentic maturity can be graded using customer signals, such as:

• When something breaks (button fails, payments fail), do you kick off incident investigation?
• Can you fix autonomously?
• What is the time/latency to resolution?
• Emphasis: shipping fixes for real product issues, not only collecting feedback.

5) Backend-as-an-agent problem (example: Supabase as an “agent-first” backend)

• A practical pattern described: always using Supabase as a backend.
• Claim: Supabase simplifies common backend tasks (DB, login, storage) and is “agent-first”, including:
  • integration where agents can create tables
  • deploy functions
  • review security within the project.

6) Execution model: minimal, fast, composable tooling (FX / “shell in a shell”)

• The stack uses CLI-first workflows, arguing that bloated agents hurt productivity due to:
  • startup latency
  • added complexity
• “FX” (Versel Labs coding CLI/harness):
  • designed to feel like a shell (fast boot, minimal binary size)
  • built to be small, embeddable, and extensible
• Core philosophy:
  • follow Unix composability
  • keep agent tooling minimal
  • extend via open interfaces rather than ever-growing UI/skills.

7) Versel “drives” and sandboxed file-system-as-interface

• Highlights file-system driven workflows:
  • the model operates with a mounted/virtual cloud drive
  • agents can read/write files for repeatability and persistence
• Values “Control O” / full traceability, including inspecting inputs/outputs of agents.

8) “Verification engineering” and guardrails over unit-tests-only thinking

• Critiques over-trusting simplistic unit tests and allowing agents to write arbitrary code/test assertions.
• Proposes alternatives:
  • guard rails
  • proof systems / correctness axioms
  • human-defined constraints and explicit tradeoffs (e.g., speed vs safety vs binary size)
• Notes systems/performance knowledge as part of effective agent use.

9) Evaluation of models + cyber benchmarks

• Mentions running models through Deepseek cybersecurity benchmarks for capability and cost.
• References Deepseek Bench / deepseec bench as described:
  • “security harnesses” can invoke coding agents/tools to uncover vulnerabilities.

---

Cybersecurity analysis and concrete examples

10) Why “code-by-LLM” increases vulnerability discovery risk

• Claims the vulnerable surface area in human code is effectively “immeasurable,” and even small code can hide large exploit space.
• Real-world example:
  • lib hif (a C/C++ image codec library) with an RCE vulnerability
  • described as widely used across the ecosystem (examples named: Slack, Next.js, OpenAI, Discourse)
• Warning:
  • cheap, strong security-reasoning models let attackers find vulnerabilities at scale.

11) “Deepseec” as defensive automation + secure-code-by-construction direction

• Advice:
  • proactively run security tooling against codebases (Deepseec and similar tools)
  • shift toward languages/libraries with stronger safety properties (e.g., memory-safe Rust) and stronger assertions.

12) Open source vs “reinventing security”

• For core infrastructure (e.g., TLS), argues for relying on libraries with higher collective scrutiny/auditing rather than building from scratch.
• Also flags uncertainty and cost/benefit tradeoffs around when to regenerate vs reuse.

13) Cloud agents and sandboxes as security and efficiency improvements

• Argues moving agents to cloud sandboxes is inevitable:
  • better compartmentalization reduces blast radius
  • easier monitoring/firewalls and containment of supply-chain risk
  • parallelization increases throughput (work can run while the user is asleep)

---

Agent infrastructure products mentioned

14) Cloud Room (open-source platform for cloud agents)

• Describes Cloud Room:
  • “world’s first open source solution for cloud agents”
  • goal: let users run agents with their own favorite models/harnesses rather than being locked into a single vendor ecosystem
• Status: not fully open to the public; waitlist at cloudroom.dev

15) AI Gateway and token spend transparency tooling

• Mentions Versel’s AI Gateway:
  • encourages investing in an “agentic engineering fabric”
  • avoid relying on only one model
• Internal tooling concept:
  • engineers use a menu/utility that shows token spend by model/provider (e.g., OpenAI vs Claude vs “Opus”)
  • encourages experimentation and observability.

---

“Main speakers / sources” (as referenced)

• Gilmore / Gilmo (speaker; CEO of Vercel; referenced repeatedly as “Gilmo”)
• Interviewer (host; referenced by dialogue cues such as “So, Gilmo…”)
• Brands/products referenced as sources/examples:
  • Vercel
  • Supabase
  • Deepseek / Deepseec benchmark
  • V0 / Vercel V0-like product
  • OpenAI / Claude (mentioned as model/tools)
  • Jeff Dean’s “Latency Numbers” resource