Evolve

Evolve is the way you change an agent in Eve Studio: you describe what you want in plain language, and Studio turns it into a real edit to the agent's files. It classifies your request, drafts the actual change with the agent's own model (its Vercel AI Gateway), shows you an exact diff, and on your approval writes the files and git-commits them. Nothing changes without an approved diff.

Why Studio does this, not the agent

An Eve agent can't edit its own files at runtime. It's filesystem-first, compiled into an immutable deployment, and read-only in production — there's no writable source tree for a live agent to reach into.

So the control plane does the authoring. Studio drafts the change, you approve it, and Studio writes it to disk and commits it. The agent stays pure base Eve — nothing bespoke wedged into the runtime — which means it stays upgradable from the base release. And because every change lands as a git commit, every change is a revert point: if an edit doesn't pan out, roll the commit back.

What a request becomes

Studio routes your intent to the right kind of capability. The same sentence can become very different files depending on what it's really asking for:

A stable behavior or persona change → instructions.md, the system prompt.
A reusable procedure → a skill.
A typed action the agent can take → a tool.
Recurring proactive work → a schedule. Local time in a schedule is converted to a UTC cron using your computer's timezone.
A durable fact about you → memory, written to the agent's Arcana brain rather than its source. No rebuild. This route needs a brain — connect one and your agent starts remembering what you tell it instead of only changing what it does.

You don't have to know which bucket you're asking for — classification is the first thing Evolve does, and the drafted diff shows you where it landed.

Three ways to evolve

1. The Evolve tab

Open an agent and go to the Evolve tab. Type an intent — for example, "Create a skill that drafts my weekly update" — and Studio drafts the change, shows the diff, and waits for your approval.

The tab also has a Scan for patterns button. It reads your recent chats and suggests recurring tasks worth turning into a skill, a tool, or a schedule — the things you keep doing by hand that the agent could own.

Connect an Arcana brain and the scan reads its timeline too, so it learns from everything the agent has actually done — including work it did on Slack or on a schedule, hours after you closed the app — not just what you typed in Studio.

2. In Studio chat

You don't have to switch tabs. When a message in chat looks like a self-change — "remind me…", "from now on…", "create a schedule…" — Studio detects it and opens the approval flow automatically. If it guessed wrong and you just wanted to talk, there's a send as a normal message escape.

3. From Slack (or any deployed channel)

You can let the agent propose its own changes from wherever it's deployed. In the Evolve tab, turn on "Let the agent propose its own changes" and redeploy. The agent gains a propose_change tool.

When you DM the agent a self-change request, it can't write files — deployed means read-only — so instead it queues the proposal. Back in Studio, the Proposals inbox at the top of the Evolve tab shows what's pending. Hit Review and it runs the same draft → diff → approve pipeline as everything else. Dismiss turns a proposal down without applying it.

Where proposals queue

The agent needs somewhere durable to put a proposal while Studio isn't watching. Studio picks the backend for you:

| | | | --- | --- | | Default | A private Vercel Blob store on the agent's own project. If there isn't one yet, the Evolve tab offers a button to create it — no keys to manage, and the token never leaves your project. | | With a brain | The agent's Arcana brain. Proposals land alongside everything else it remembers, so the same store holds its memory and its pending changes. |

Either way the inbox works the same. A brain also unlocks the memory route and gives Scan for patterns a lot more to learn from.

Review and apply

Every route ends at the same gate: a diff you approve.

Before you approve, you can Edit the drafted change — tweak the wording, rename a file, adjust the cron. When you're happy, approve, and Studio writes the files and commits.

To pick up the change:

Restart the agent to load it locally, or
Deploy for production.