Skip to content
Managed Agents

Independent proposal · Built in seven days, in the open

Agent teams from your repository or your idea

Point it at a codebase or describe the job. Get agents you can see into, steer and stop, with every action, memory and cost on the record.

See how it worksRead the changelog
Relay workspaceSample data

Engineering Review

Fix the invitation flow

relay/web at 3f2c1a9

  1. Inspected the repository: Done

    14 files read, no scripts run

  2. Reproduced the bug: Done

    Expired invitations still accepted

  3. Patch and regression tests: Running

    4 cases: valid, expired, unauthorized, duplicate

  4. Post the review: Needs approval

    Waiting for your approval

Approval needed

Action
Post review comment
Target
relay/web pull request 128
Effect
1 comment, visible to the repository's members
Model: Claude via AI Gateway, GPT allowed as fallbackCost: $0.42 estimated, reconciles after the run

01 / How it works

One lifecycle, from first look to retirement

The same agent, version and permissions follow the work through every step, whether it started from a repository or from an idea.

Understand the project before proposing anything

Sample

Read the files you allow, without running the project's scripts. Every recommendation separates what the code shows from what we only suspect, and says what access it would need.

relay/web at 3f2c1a9 · 14 files read

Invitations expire after 7 daysObserved
Support answers invite questions by handHypothesis

Suggested: invitation helper

For customers · needs read access to invitations · no write access

02 / Teams

Senior people lead. Agents do the legwork.

Four teams at Relay, a fictional company, each led by a person who owns the goal and the permissions. These are the demos the product is being built to pass.

  • Go to market

    Led by the VP of GTM

    A launch plan for 20 target accounts, with campaign drafts ready for review.

    • Segment analyst
    • Account researcher
    • Campaign planner
    • Enablement reviewer

    Sends nothing. Every outbound message stays a draft.

  • Engineering

    Led by the staff engineer

    A failing invitation flow repaired, tested and ready for review.

    • Diagnosis
    • Implementation
    • Tests
    • Instrumentation
    • Security

    No merges, no production secrets, no production deploys.

  • Design

    Led by the design director

    Two onboarding directions to critique, then one ready to build.

    • Research synthesis
    • Interaction
    • Systems
    • Prototyping
    • Accessibility review

    Shared design files change only with a grant.

  • Product marketing

    Led by the head of product marketing

    Positioning and claims other teams can reuse with confidence.

    • Insight analyst
    • Competitive research
    • Positioning
    • Claims audit
    • Enablement writing

    Unsupported claims are blocked. Nothing is published.

03 / Control

You stay in charge while the work runs

The product makes the consequential parts visible: who can do what, what was remembered, what it cost and what happens when you stop it.

  • Coworkers, not copies

    Adding a shared agent gives you access to the same one your team uses. Ten people asking for the same review get one review, and private chats stay private.

  • Approval for the exact action

    You approve a specific action on a specific target. If either changes, it asks again.

  • Memory with receipts

    See where each memory came from and who can use it. Correct it or forget it.

  • Cost on the record

    Estimated while it runs, reconciled after. Budget limits stop new work instead of quietly spending more.

  • Models you can swap

    Choose between model families through one gateway. Fallbacks stay inside the privacy rules you set, and every run shows the model that answered.

04 / Built on

Public building blocks, our own product

The proposal builds on these technologies. They are named as references only.

Eve
Agent framework, beta, pinned behind an adapter
AI Gateway
One route to several model families
AI SDK
Streaming, tools and structured output
Postgres
Durable state, approvals and the effect ledger
Next.js
The workspace and this site

05 / Changelog

Latest changes

All changes

06 / Questions

Frequently asked

Is this a Vercel product?

No. Independent proposal; not affiliated with or endorsed by Vercel. It is a proposal for how a managed home for agents could work, built on public technology.

What is real and what is sample data?

Screens on this page use sample data for Relay, a fictional company, and are labeled. The changelog records what actually works in the deployed preview and what is still blocked.

Why is the name marked as a working name?

Managed Agents is a neutral placeholder until the product has a name of its own.

Which models does it use?

At least two model families through one gateway, chosen from the gateway's catalog when they are evaluated. Each run shows which model actually answered.

How do I follow progress?

The changelog has one entry per day of the build, with evidence and open blockers, and an RSS feed.