Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Give Your Agent a Computer — Nico Albanese, Vercel
TL;DR
Nico’s core thesis is simple: real agents need three things — a runtime, tools, and a computer — he argues the missing piece in 2026 agent design is giving the model a sandboxed file system so it can persist state, run code, and stay on task across long sessions.
AI SDK 6 is pushing agents toward reusable, typed definitions instead of giant 2,000-line route handlers — Nico shows how
toolLoopAgent, global providers via the AI Gateway, and end-to-end typed messages let you define an agent once inlib/agent.tsand use it cleanly across Next.js apps and beyond.Provider-executed tools like OpenAI web search are powerful because they require almost no code, but the UX falls apart unless you render tool activity in the UI — his demo adds
webSearchPreviewin minutes, then usesinferAgentUIMessageso the tool inputs and outputs become type-safe all the way to the React client.The big unlock came internally when Vercel agents got a scratchpad file system — Nico says an internal data agent went from vaguely hallucinated answers using 5–10 tools to reliably finishing tasks once each session had a plan file, research directory, and explicit instructions to check work off step by step.
Vercel’s new persistent sandbox model treats a named computer like durable infrastructure, not a disposable container — instead of manually snapshotting and restoring environments, you reference a sandbox by name and Vercel spins it up, pauses it after inactivity, then restores the same filesystem state on the next request.
Nico is skeptical that context compaction is the main bottleneck now that models have million-token windows — he cites his own coding agent running for 104 minutes, making 316 tool calls, changing 29 files, using only 32% of GPT-5.4’s context window, while maintaining a 95% cache-token read ratio and leaning more on sub-agents than summarization.
The Breakdown
Bootstrapping the workshop: clone, link, install, go
Nico opens by live-building the demo from scratch: clone the GitHub repo, install the Vercel CLI, run vercel link, and pull env vars to get the OIDC token needed for both AI Gateway inference and Vercel Sandbox auth. It’s very hands-on and lightly chaotic in a charming way — package-manager confusion, outdated CLI auth, and shoutouts to someone’s homemade TypeScript shirt all make it feel like a real workshop, not a polished launch reel.
From “four primitives” to a reusable agent definition
He frames AI SDK’s evolution from the older generateText / streamText era to a more object-oriented agent model built to avoid the dreaded chat.ts file with 2,000 lines of inline tools and prompts. The demo agent is almost hilariously small: a toolLoopAgent in lib/agent.ts using gpt-5.4-mini, with AI SDK 6’s global provider concept making it possible to call models through plain strings via the AI Gateway.
Wiring the UI and proving prompts still matter
Next comes the route handler and the familiar useChat hook, which Nico calls the original AI SDK component from three years ago. Once the basic chatbot works, he changes the instructions to “respond like a cowboy” and gets “Howdy partner,” using that tiny example to make a bigger point: system prompts may feel old-school, but they’re still one of the core control surfaces for agents.
Adding web search, then fixing the invisible-tool problem
To give the agent more context, he adds OpenAI’s provider-executed web search tool, walking through his taxonomy of custom tools, provider-defined tools, and provider-executed tools. The first result works, but the user sees only an awkward pause — so he pivots to rendering tool states in the UI and shows off one of AI SDK’s strongest ideas: the agent definition should be the source of truth, with typed messages flowing end-to-end via inferAgentUIMessage.
Why a file system changed everything inside Vercel
This is the philosophical center of the talk. Nico describes Vercel’s internal agents — especially a data-focused Slackbot replacement called DZero — and says performance jumped when they got a scratchpad: a plan file with the objective at the top, research folders, and persistent artifacts instead of stuffing everything into an ever-growing context window. He says support agents built this way cut support tickets by about 90%, with around 95% of users actually replying “thank you, thank you.”
Persistent sandboxes: giving an agent “its own computer”
He introduces Vercel Sandbox beta and, specifically, named persistent sandboxes — a cleaner answer to the usual ephemerality of cloud sandboxes. The pitch is that your app references a sandbox by name, Vercel routes to an active session if one exists, or spins up a new one and restores the snapshot, which gives the agent something that feels like a stable personal machine without ugly lifecycle-management code.
Bash as the universal tool, plus a detour into context strategy
The actual implementation is almost aggressively minimal: define a bash tool, inject the sandbox into runtime context through typed call options, and let the agent run shell commands. Nico pauses for a long, opinionated answer on memory and context pruning, arguing that compaction can break cache efficiency and even erase critical instructions; his own coding agent, he says, ran for 104 minutes with 316 tool calls and 29 files changed while using just 32% of GPT-5.4’s context, so he’s leaning harder into sub-agents than aggressive summarization.
Memory as a file, and the first glimpse of self-extension
The last stretch is the fun part: he stores memory in memories.md, injects it into the system prompt on each run, and has the agent update it with durable facts like “user’s name is Nico” and “works on the AI SDK at Vercel.” Then he pushes the idea further: if a task repeats, the agent should write a Python script, save it, and remember that it exists — so when asked for weather twice, it starts building a little toolbox for itself inside its own sandbox.
The bigger system behind the demo
Nico closes by showing a more serious internal-style coding agent built on the same patterns: AI SDK, AI Gateway, durable workflow steps, resumable streams, and sub-agents that do exploratory work off the main thread. It’s already being used by 23 people at Vercel, has processed about 3.8 billion tokens in a month or two, sits around a 91% cache read ratio, and has been responsible for nearly 350 PRs — his way of saying this isn’t just a workshop trick, it scales.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
