Amvy: How multiplayer AI agents work for teams

One session. Your team. Your agents.

Stop pasting agent transcripts into DMs. In AMVY, multiplayer AI means people and agents share one governed session: everyone sees the same work, and anyone with the right role can step in.

Release readiness · Shared

3 online

Ana Pérez10:42

@AMVY Review open issues, risky changes, and ship blockers for the 3.2 release.

AMVY

Working for 48s

8 pull requests reviewed against release policy·acme/platform

Checking required release checks·release-policy.md

Ana Pérez is working with the agent

Active now

Built on the Multiplayer AI Manifesto

Read the manifesto

Never copy-and-paste
One canonical session that everyone joins.
Work with the door open
Watch, steer, and take over live runs.
Continuously improve
Reviewed corrections become skills and memory.
People are not routers
Agents and schedules move the work along.
Nothing starts from scratch
Threads, workspaces, and history resume.

One request. A complete operating loop.

AMVY carries context, execution state, approvals, and results across the entire job — not just one response.

Start with the outcome

Describe the work in natural language and attach the project, files, or systems that matter.

Project · Platform release

GitHub · 8 open pull requests

Goal · Produce a readiness report

Work moves without anyone chasing it.

People are not routers. Agents run on a schedule or a trigger, split long work across subagents, and ask the right person when they need an answer.

Schedules
Recurring or one-time runs for a person or a whole project, with the same approvals as any other run.
Triggers
Start a run from a webhook or the API when something happens in your systems.
Background work
Long jobs keep running after you close the tab, and subagents split the work in parallel.
Agents ask people
When a run needs a fact only a teammate knows, the agent asks them in their Inbox and continues with the answer.

Operations daily · Shared

2 online

Schedule · every weekday09:00

Check overdue tasks across our projects and chase the owners.

AMVY

Worked for 3m 40s

9 overdue tasks found·Asana

4 projects checked in parallel·4 subagents

Asked 3 owners for an update·Microsoft Teams

Question from AMVY · for Laura Díaz

Can the onboarding checklist move to Friday? It is blocked on the new vendor contract.

Asked during the daily operations check started by Tomás Vidal

Your agents know what your company knows.

Answers come from your documents, the folders you already keep, and the facts your teams agreed on, never from a guess.

Knowledge bases
Upload playbooks, policies, and specs, and link them to an agent or a project.
Connected sources
Folders and sites stay in sync on their own, with the permissions they already have.
Organization memory
Facts your teams agree on become shared memory, reviewed before any agent uses them.
Scoped to each person
An agent only reads what the person and the project it works for can read.

SharePointOneDrivewebsites and the files you upload

Acme renewal

1 online

Marta Díaz14:05

@AMVY What discount can we offer Acme on the annual plan?

AMVY

Worked for 21s

Searched the sales playbook·Knowledge base

Read Pricing 2026.xlsx·SharePoint

Recalled 2 organization memories·Organization memory

Up to 15% on annual plans; anything above needs finance approval. Acme already has 10% from last year's renewal, so you can add 5% more within policy.

Organization memory

Annual discounts above 15% need finance approvalReviewed
Acme renewed in 2025 with a 10% discountReviewed

It learns through proposals, not silent changes.

Self-improving agents, under review. Completed work can produce evidence-backed memory or skill proposals, and people approve, edit, or reject every durable change.

Run completed
Learning proposed
Human reviewed
Memory or skill updated

Skill patchPending review

Add a rollback check to Release review

The last run needed rollback steps that the skill did not cover.

ScopeProject

Targetrelease-review

PermissionUpdate skills

Confidence86%

Pathskills/release-review/SKILL.md

Proposed change

Skill patch

Current

  ## Checks  - Verify CI is green  - Review open blockers

Proposed

  ## Checks  - Verify CI is green  - Review open blockers+ - Confirm a tested rollback plan exists+ - Name the rollback owner in the report

Why this was proposed

The release run had to add a rollback plan by hand. Adding it to the skill means the next run starts with it.

How often to suggest

Pick a cadence. Reviews stay within quotas and the Inbox stays quiet.

Up to 3 proposals per run and 10 per day.

Inbox only by default: agents propose, people apply.
Permissions are checked again when a change is applied.
Applied, rejected, and failed proposals stay auditable.

A real workspace for every run.

Project files, diffs, tools, artifacts, and execution history stay together in a persistent environment scoped to the person or project.

Files & diffs

Complete run history

Isolated runtime

Persistent state

Dedicated project runtime

workspace

src/release/checks.ts

reports/readiness.md

artifacts/release-room/

skills/release-review/

Run activity · live

Project context and release policy loaded

Analyzing 8 pull requests and 42 checks...

All required checks passed

4 files changed · 28 insertions · 6 deletions

Run completed · report generated

Artifact published/a/release-control-roompublished

Release control room

published

/a/release-control-room · v3 · Organization

Don’t stop at an answer. Publish the result.

Turn completed work into private or organization-visible dashboards, reports, portals, and interactive tools with versioned publishing.

Interactive dashboards

Decision reports

Internal portals

Operational trackers

Know an agent works before your team relies on it.

Turn real conversations into test cases and replay them whenever an agent, a skill, or a model changes.

Cases from real work
Save any conversation as a case with the outcome you expect.
No real side effects
Replays run with stubbed tools in a sandbox without network, so nothing is sent or changed.
Assertions and an AI judge
Check the exact steps you require and score the answer against your criteria, within a budget per run.
Every change gets tested
Suites run nightly, and a self-improvement proposal is compared against the current setup before anyone applies it.