Summary

Vercel is turning durable workflows into an everyday primitive for production agents. Its open-source Workflow Development Kit lets code pause for days at no compute cost, retry failed steps, resume from webhooks, and support human approval while still reading like ordinary application code. Malte Ubl’s ambition is to make workflows “something you do literally every day,” not an architecture requiring 15 design documents. Malte also described Vercel’s open-source model as growing the overall pie while retaining a roughly proportional share, and expects some users to run the software themselves.
AI SDK 6 adds an agent abstraction only after repeated usage patterns began to stabilize. Vercel deliberately kept earlier versions low-level because, unlike mature web development, “we still know absolutely nothing” about the eventual shape of AI applications; the abstraction already appeared experimentally in AI SDK 5. Malte accepts that models need freedom to control tool calls, but argues the overlooked platform problem is embedding that behavior into applications with streaming, controls, and application integration.
Vercel’s strongest agent advantage may be application context that standalone tools do not have together. Vercel Agent can see runtime data, logs, deployments, how to start the development server, and existing secrets. Its DevOps agent combines conventional anomaly detection with an investigative LLM, addressing the recall-precision trade-off by aggressively flagging anomalies, allowing the agent to investigate for around two minutes, and waking a human only when escalation appears warranted.
The company is drawing a hard line between narrow agent competence and unsafe autonomy. Malte says agents are simultaneously “extraordinarily effective and still very ineffective”: qualifying leads, preparing abuse investigations, and querying well-documented warehouse semantics work; changing a firewall or performing DNS migrations does not yet belong in the autonomy bucket. The operating heuristic is simple: ask employees, “What do you hate most about your job?”
“Agent on Every Desk” is both an enterprise adoption motion and a product-learning loop. Vercel asks participating companies to commit to three agents: Vercel builds the first, supports the customer closely on the second, and expects the customer to build the third independently. Startups should instead take the open-source examples, feed them to a coding agent, and adapt them without a forward-deployed team.
Vercel is broadening its platform to Python alongside TypeScript. It shipped zero-config Flask and FastAPI deployment plus a Python SDK, with Fluid Compute charging only for active CPU time—a material fit when an AI backend takes 30 seconds to respond. Malte expects PHP and Ruby eventually, but says native developer experience, not VM-level feasibility, is the limiting work.
AI-assisted coding forces infrastructure to distrust both the developer and the generated application. Vercel wants systems that remain secure even when “the developer doesn’t know what they’re doing” and the AI does not either, moving authentication and data-access enforcement outside application-controlled code. Strategically, this turns safe deployment—not merely code generation—into a critical layer of the agent-native stack.

Deep dive

1. Durable workflows make long-running agents ordinary application code

Malte’s conference-level framing was deliberately concrete: agents are exciting because “you can actually build them.” Vercel builds customer products and internal operational agents first, then extracts abstractions grounded in those applications rather than designing from theory.
The Workflow Development Kit makes durable, resumable, streamable execution feel idiomatic. A function can effectively run indefinitely, pause between steps without consuming compute, resume days later, and automatically retry a failed step instead of forcing developers to assemble queues, state tables, and recovery cron jobs.
The host’s useful mental model was “an indefinitely long-running” serverless function. For human approval, a workflow creates an ephemeral webhook, stores its URL, awaits it for free, and continues when a user approves the item two hours—or potentially much longer—later.
The host noted that production users want ownership, auditability, and the option to self-host; Malte agreed that people will run the software themselves. He framed Vercel’s open-source model against support-only and open-core approaches: the software is genuinely runnable, which grows the overall pie while keeping Vercel’s share roughly proportional.

2. AI SDK stayed thin until agent patterns earned an abstraction

AI SDK 6 beta promotes a direct agent abstraction that existed experimentally in AI SDK 5. Previously, developers built an agent as a tool-equipped streamText loop, with controls for preparing each step and selecting tools; competing frameworks could—and often did—build thicker abstractions on top.
Malte’s explanation for that restraint: a 2025 web-framework author already knows what users will build, while AI applications remain radically unsettled. “We know absolutely nothing, and we still know absolutely nothing,” so a thick, prescriptive abstraction is unusually likely to freeze the wrong assumptions.
The host pushed back that frontier-model labs intentionally prefer a Hollywood-principle framework—give the wheel to the model and let it choose tools and reasoning. Malte agreed that developers must “let go and let the model kind of take control,” especially for coding agents, but separated model autonomy from the application-layer problem of embedding and operating it.
Streaming illustrates that application-layer value: web developers once abandoned features over 500 milliseconds, while model responses can take 30 seconds. Vercel focused on making streaming intuitive, then validated its abstractions through dogfooding—AI SDK came from v0, and Vercel later did the substantial work of moving v0 fully back onto it.

3. Application context turns anomaly alerts into agent investigations

Malte distinguished internal custom agents from Vercel Agent, an “agent as a service” product. Its edge is integrated context: Vercel already has runtime telemetry, error logs, deployment data, dev-server startup knowledge, and secrets, avoiding the experience of onboarding an external agent “like onboarding a junior employee.”
The newly announced DevOps agent connects to Vercel’s anomaly-detection system. When production behavior deviates, the agent can build and execute observability queries, inspect time-series slices, drill into logs, and present both its diagnosis and the graphs it used; Malte said it “almost all the time” identifies what happened very precisely.
The host’s pushback—worth keeping—was that language models are not naturally time-series models, alert preferences differ, and SRE teams risk “always fighting the last war.” Malte clarified that the LLM does not detect anomalies: a separate pipeline over Vercel’s time-series database does that, after which the agent investigates with query tools and text logs.
That division improves the classic recall-precision trade-off. Detection can be tuned aggressively without paging someone for a harmless newsletter-driven traffic spike; a “co-worker that has no sleeping problems” can spend two minutes investigating and decide whether to escalate. Future source-code access might enable occasional repair PRs, but Malte would not yet let it modify firewalls—or migrate DNS.

4. The best internal agents remove hated work without removing judgment

Malte’s discovery question is “What do you hate most about your job?” The answers tend to be repetitive, text-heavy tasks requiring enough judgment to resist conventional automation, yet consuming enough employee time to carry meaningful business impact.
Vercel’s lead-qualification agent researches incoming contact-sales requests through tools such as LinkedIn and Google, applies the company’s qualification criteria, and redirects misclassified support requests. Malte treated this as a comparatively straightforward, soft-case agent use case and open-sourced Vercel’s implementation.
Its abuse-analysis agent performs pre-work rather than making the final decision: it visits the reported site, checks account age and billing history, and can recognize something resembling a Facebook login page. A human still reviews the assembled evidence and determines the enforcement action.
The open-source data-analyst agent reflects a more qualified claim. Vercel does not promise magical prompt-to-SQL from schema alone; it created a structured method for documenting warehouse semantics, supplying the context needed for the agent to answer internal questions successfully enough for regular use.

5. Forward deployment teaches enterprises to build the second and third agent

“Agent on Every Desk” targets large companies with substantial potential efficiencies but no confidence choosing or shipping a first project. Forward-deployed engineers help identify work that is both high-impact and feasible, then implement the initial agent inside the customer’s environment.
The contract encodes capability transfer: customers commit to three agents; Vercel builds the first, stays alongside them for the second, and assumes the third will need little help. “We don’t want to stay there”—the end state is a company able to build its own custom agents.
Malte would not prescribe that model to startups. His alternative is the open-source artifact plus a coding agent: show Claude Code the example, describe the company-specific differences, and build from there. Enterprise engagements, meanwhile, extend Vercel’s dogfooding loop into problems a “500-person startup” might never encounter alone.

6. Python support and security guardrails broaden the AI platform

Vercel shipped zero-config support for Flask, FastAPI, and other popular Python frameworks, plus a Python SDK for its API. On Fluid Compute, active-CPU pricing means a Python service pays while computing but not while idle—particularly relevant when an AI backend needs 30 seconds to respond.
Malte separated Fluid Compute from durable workflows: between workflow steps “literally nothing’s running,” while Fluid keeps a VM available for more agile work. TypeScript had become GitHub’s largest language “as of today,” he noted, but Vercel sees both ecosystems as large and relevant; PHP and Ruby should follow once their developer experience feels native.
His broader leadership lesson came from transforming a pre-ChatGPT company through bets that felt native: v0 began as a web-page creation tool, while AI SDK extended Vercel’s identity as a framework company. He also defended strong individual-contributor ladders so excellent engineers need not choose between stalled compensation and becoming “a potentially very bad manager.”
As designers, product managers, and other employees contribute AI-generated code, Vercel is designing for a threat model where neither developer nor model is assumed competent. Authentication and authorization over visible data must sit outside app-controlled logic, providing “a minimum amount of security fully independent of the quality of the app.”

Sarah Paine — How Russia sabotaged China's rise Marc Andreessen and Ben Horowitz on the State of AI